The test started failing after my commit:
ca07088c
(func: fix not unloading of unused modules), because I forgot to
update the result file.

Follow up #4648

This patch fixes a bug that prevented the conversion of real
values that are greater than INT64_MAX and less than UINT64_MAX to
INTEGER and UNSIGNED.

Closes #4526

C functions are loaded from .so/.dylib dynamic libraries. A
library is loaded when any function from there is called first
time. And was supposed to be unloaded, when all its functions are
dropped from the schema (box.schema.func.drop()), and none of them
is still in a call. But the unloading part was broken.

In fact, box.schema.func.drop() never unloaded anything. Moreover,
when functions from the module were added again without a restart,
it led to a second mmap of the same module. And so on, the same
library could be loaded any number of times.

The problem was in a useless flag in struct module preventing its
unloading even when it is totally unused. It is dropped.

Closes #4648

Error injections are used to simulate an error. They are
represented as a flag, or a number, and are used in Lua tests. But
they don't have any feedback. That makes impossible to use the
injections to check that something has happened. Something very
needed to be checked, and impossible to check in a different way.

More certainly, the patch is motivated by a necessity to count
loaded dynamic libraries to ensure, that they are loaded and
unloaded when expected. This is impossible to do in a platform
independent way. But an error injection as a debug-only counter
would solve the problem.

Needed for #4648

There is a bug in XCode 11 which makes some standard C headers
not self sufficient when compile with gcc. At least <stdlib.h> and
<algorithm> are affected. When they are included first,
compilation fails with creepy errors like this:

    In file included
    from /Applications/Xcode.app/Contents/Developer/
        Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/
        sys/wait.h:110,
    from /Applications/Xcode.app/Contents/Developer/
        Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/
        stdlib.h:66,
    from tarantool/third_party/zstd/lib/common/zstd_common.c:16:
        /Applications/Xcode.app/Content/Developer/
        Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/
        sys/resource.h:
    In function 'getiopolicy_np': /Applications/Xcode.app/Contents/Developer/
        Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/
        sys/resource.h:447:34: error:
            expected declaration specifiers before '__OSX_AVAILABLE_STARTING'
            447 | int     getiopolicy_np(int, int)
            __OSX_AVAILABLE_STARTING(__MAC_10_5, __IPHONE_2_0);

The patch workarounds the bug by deleting the buggy header
includes where possible, and by changing include order in other
cases.

Also there was a second compilation problem. This was about
different definitions of the same standard functions: via extern
"C" and without. It looked like this:

    In file included from tarantool/src/trivia/util.h:36,
    from tarantool/src/tt_pthread.h:35,
    from tarantool/src/lib/core/fiber.h:38,
    from tarantool/src/lib/core/coio.h:33,
    from tarantool/src/lib/core/coio.cc:31:
    /usr/local/Cellar/gcc/9.2.0_1/lib/gcc/9/gcc/x86_64-apple-darwin18/9.2.0
    include-fixed/stdio.h:222:7: error: conflicting declaration of
    'char* ctermid(char*)' with 'C' linkage
      222 | char *ctermid(char *);
          |       ^~~~~~~

    In file included from /Applications/Xcode.app/Contents/Developer/Platforms/
    MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/unistd.h:525,
    from tarantool/src/lib/core/fiber.h:37,
    from tarantool/src/lib/core/coio.h:33,
    from tarantool/src/lib/core/coio.cc:31:
    /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/
    Developer/SDKs/MacOSX.sdk/usr/include/_ctermid.h:26:10: note: previous
    declaration with 'C++' linkage
      26 | char    *ctermid(char *);
         |          ^~~~~~~

This bug is workarounded by deletion of the conflicting includes,
because anyway they appeared to be not needed.

Closes #4580

Conflicts:
	third_party/decNumber

Unicode_ci collation breaks the general
rule for objects naming, so we remove it
in version 2.3.1

Closes #4561

The problem was in that the test uses the global trigger
box.session.on_disconnect() to set a global variable by one
connection. But test-run can do multiple connects/reconnects to
the same instance. That led to multiple invocations of
box.session.on_disconnect(), which could override the global
variable in unexpected ways and moments.

The patch makes only one session execute that trigger.

Probably related to https://github.com/tarantool/test-run/issues/46


Follow up #4627

Reviewed-by: Alexander Turenko <alexander.turenko@tarantool.org>

There were some pass conditions in quorum test which could take some
time to be satisfied. Now they are wrapped using test_run:wait_cond to
make the test stable.

Closes #4586

In replicaset_follow we iterate anon replicas list: list of replicas
that haven't received an UUID. In case of successful connect replica
link is being removed from anon list. If it happens immediately,
without yield in applier, iteration breaks. Now it is fixed by
rlist_foreach_entry_safe instead of common rlist_foreach_entry.
Relevant test case is added.

Part of #4586
Closes #4576
Closes #4440

Binary session disconnect trigger yield could lead to use after
free of the session object. That happened because iproto thread
sent two requests to TX thread at disconnect:

    - Close the session and run its on disconnect triggers;

    - If all requests are handled, destroy the session.

When a connection is idle, all requests are handled, so both these
requests are sent. If the first one yielded in TX thread, the
second one arrived and destroyed the session right under the feet
of the first one.

This can be solved in two ways - in TX thread, and in iproto
thread.

Iproto thread solution (which is chosen in the patch): just don't
send destroy request until disconnect returns back to iproto
thread.

TX thread solution (alternative): add a flag which says whether
disconnect is processed by TX. When destroy request arrives, it
checks the flag. If disconnect is not done, the destroy request
waits on a condition variable until it is.

The iproto is a bit tricker to implement, but it looks more
correct.

Closes #4627

Replication's applier encoded an auth request with exactly the
same parameters as extracted by the URI parser. I.e. when no
password was specified, the parser returned it as NULL, and it was
not encoded. The relay, received such an auth request, complained
that IPROTO_TUPLE field is not specified (this is password).

Such an error confuses - a user didn't do anything illegal, he
just used URI like 'login@host:port', without a password after the
login.

The patch makes the applier use an empty string as a default
password.

An alternative was to force a user always set a password even if
it is an empty string, like that: 'login:@host:port'. And if a
password was not found in an auth request, then reject it with a
password mismatch error. But in that case a URI of kind
'login@host:port' becomes useless - it can never pass. In
addition, netbox already uses an empty string as a default
password. So the only way to make it consistent, and don't break
anything - repeat netbox logic for replication URIs.

Closes #4605

Conflicts:
	test/replication/suite.cfg

Box.info.replication shows applier/relay's latest error message.
But it didn't include errno description for system errors, even
though it was included in the logs. Now box.info shows the errno
description as well, when possible.

Closes #4402

Conflicts:
	test/replication/suite.cfg

The only error type having an errno as a part of it was
SystemError (and its descendants SocketError, TimedOut, OOM, ...).
That was used in logs (SystemError::log() method), and exposed to
Lua (if type was SystemError, an error object had 'errno' field).

But actually errno might be useful not only there. For example,
box.info.replication exposes the latest error message of
applier/relay as 'message' field of 'upstream/downstream' fields,
lacking errno description.

Before the patch it was impossible to obtain an errno code from C,
because it was necessary to check whether an error has SystemError
type, cast to SystemError class, and call SystemError::get_errno()
method.

Now errno is available as a part of struct error object (available
from C), and is not 0 for system errors.

Part of #4402

Box.session.su() raised 'SystemError' when a user was not found
due to a too long user name. That was obviously wrong, because
SystemError is always something related to libraries (standard,
curl, etc), and it has an errno code.

Now a ClientError is raised.

fiber.top() fills in statistics every event loop iteration,
so if it was just enabled, fiber.top() returns zero in fiber cpu
usage statistics because total time consumed by the main thread was
not yet accounted for.
Same stands for viewing top() results for a freshly created fiber:
its metrics will be zero since it hasn't lived a full ev loop iteration
yet.
Fix this by delaying the test till top() results are meaningful and add
minor refactoring.

Follow-up #2694

The problem appears after 6c627af3
('test: tarantoolctl: verify delayed box.cfg()'), where the test case
was changed and it doesn't more assume an error at the instance start.
So we need to stop it to prevent a situation when instances are stay
after `make test`.

Fixes #4600.
Reviewed-by: Vladislav Shpilevoy <v.shpilevoy@tarantool.org>

Before commit 03f85d4c ('app: fix
boolean handling in argparse module') the module does not expect a value
after a 'boolean' argument. However there was the problem: a 'boolean'
argument can be passed only at end of an argument list, otherwise it
wrongly consumes a next argument and gives a confusing error message.

The mentioned commit fixes this behaviour in the following way: it still
allows to pass a 'boolean' argument at end of the list w/o a value, but
requires a value ('true', 'false', '1', '0') if a 'boolean' argument is
not at the end to be provided using {'--foo=true'} or {'--foo', 'true'}
syntax.

Here this behaviour is changed: a 'boolean' argument does not assume an
explicitly passed value despite its position in an argument list. If a
'boolean' argument appears in the list, then argparse.parse() returns
`true` for its value (a list of `true` values in case of 'boolean+'
argument), otherwise it will not be added to the result.

This change also makes the behaviour of long (--foo) and short (-f)
'boolean' options consistent.

The motivation of the change is simple: it is easier and more natural to
type, say, `tarantoolctl cat --show-system 00000000000000000000.snap`
then `tarantoolctl cat --show-system true 00000000000000000000.snap`.

This commit adds several new test cases, but it does not mean that we
guarantee that the module behaviour will not be changed around some
corner cases, say, handling of 'boolean+' arguments. This is internal
module.

Follows up #4076.
Reviewed-by: Vladislav Shpilevoy <v.shpilevoy@tarantool.org>

Bootstrap and recovery work on behalf of admin. Without the
universe access they are not able to even fill system spaces with
data.

It is better to forbid this ability until someone made their
cluster unrecoverable.

The admin user has universal privileges before bootstrap or
recovery are done. That allows to, for example, bootstrap from a
remote master, because to do that the admin should be able to
insert into system spaces, such as _priv.

But after the patch on online credentials update was implemented
(#2763, 48d00b0e) the admin could
loose its universal access if, for example, a role was granted to
him before universal access was recovered.

That happened by two reasons:

    - Any change in access rights, even in granted roles, led to
      rebuild of universal access;

    - Any change in access rights updated the universal access in
      all existing sessions, thanks to #2763.

What happened: two tarantools were started. One of them master,
granted 'replication' role to admin. Second node, slave, tried to
bootstrap from the master. The slave created an admin session and
started loading data. After it loaded 'grant replication role to
admin' command, this nullified admin universal access everywhere,
including this session. Next rows could not be applied.

Closes #4606

After the issue #4537 fixed for the data segment size limit,
the temporary blocked tests because of it unblocked.

Part of #4271

Implement a new function in Lua fiber library: top(). It returns a table
containing fiber cpu usage stats. The table has two entries:
"cpu_misses" and "cpu". "cpu" itself is a table listing all the alive
fibers and their cpu consumtion.
The patch relies on CPU timestamp counter to measure each fiber's time
share.

Closes #2694

@TarantoolBot document
Title: fiber: new function `fiber.top()`

`fiber.top()` returns a table of all alive fibers and lists their cpu
consumption. Let's take a look at the example:
```
tarantool> fiber.top()
---
- cpu:
    107/lua:
      instant: 30.967324490456
      time: 0.351821993
      average: 25.582738345233
    104/lua:
      instant: 9.6473633128437
      time: 0.110869897
      average: 7.9693406131877
    101/on_shutdown:
      instant: 0
      time: 0
      average: 0
    103/lua:
      instant: 9.8026528631511
      time: 0.112641118
      average: 18.138387232255
    106/lua:
      instant: 20.071174377224
      time: 0.226901357
      average: 17.077908441831
    102/interactive:
      instant: 0
      time: 9.6858e-05
      average: 0
    105/lua:
      instant: 9.2461986412164
      time: 0.10657528
      average: 7.7068458630827
    1/sched:
      instant: 20.265286315108
      time: 0.237095335
      average: 23.141537169257
  cpu_misses: 0
...

```
The two entries in a table returned by `fiber.top()` are
`cpu_misses` and `cpu`.

`cpu` itself is a table whose keys are strings containing fiber ids and
names.
The three metrics available for each fiber are:
1) instant (per cent),
which indicates the share of time fiber was executing during the
previous event loop iteration
2) average (per cent), which is calculated as an exponential moving
average of `instant` values over all previous event loop iterations.
3) time (seconds), which estimates how much cpu time each fiber spent
processing during its lifetime.

More info on `cpu_misses` field returned by `fiber.top()`:
`cpu_misses` indicates the amount of times tx thread detected it was
rescheduled on a different cpu core during the last event loop
iteration.
fiber.top() uses cpu timestamp counter to measure each fiber's execution
time. However, each cpu core may have its own counter value (you can
only rely on counter deltas if both measurements were taken on the same
core, otherwise the delta may even get negative).
When tx thread is rescheduled to a different cpu core, tarantool just
assumes cpu delta was zero for the latest measurement. This loweres
precision of our computations, so the bigger `cpu misses` value the
lower the precision of fiber.top() results.

Fiber.top() doesn't work on arm architecture at the moment.

Please note, that enabling fiber.top() slows down fiber switching by
about 15 per cent, so it is disabled by default.
To enable it you need to issue `fiber.top_enable()`.
You can disable it back after you finished debugging  using
`fiber.top_disable()`.
"Time" entry is also added to each fibers output in fiber.info()
(it duplicates "time" entry from fiber.top().cpu per fiber).
Note, that "time" is only counted while fiber.top is enabled.

Before the patch update was implemented as a set of operations
applicable for arrays only. It was ok until field names and JSON
paths appearance, because tuple is an array on the top level.

But now there are four reasons to allow more complex updates of
tuple field internals by JSON paths:

  - tuple field access by JSON path is allowed so for consistency
    JSON paths should be allowed in updates as well;

  - JSON indexes are supported. JSON update should be able to
    change an indexed field without rewriting half of a tuple, and
    its full replacement;

  - Tarantool is going to support documents in storage so JSON
    path updates is one more step forward;

  - JSON updates are going to be faster and more compact in WAL
    than get + in-memory Lua/connector update + replace (or update
    of a whole tuple field).

The patch reworks the current update code in such a way, that now
update is not just an array of operations, applied to tuple's top
level fields. Now it is a tree, just like tuples are.

The concept is to build a tree of xrow_update_field objects. Each
updates a part of a tuple. Leafs in the tree contain update
operations, specified by a user, as xrow_update_op objects.

To make the code support and understanding simpler, the patch
splits update implementation into several independent
files-modules for each type of an updated field. One file
describes how to update an array field, another file - how to
update a map field, etc. This commit introduces only array. Just
because it was already supported before the patch. Next commits
will introduce more types one by one.

Besides, the patch makes some minor changes, not separable from
this commit:

  - The big comment about xrow updates in xrow_update.c is
    updated. Now it describes the tree-idea presented above;

  - Comments were properly aligned by 66 symbols in all the moved
    or changed code. Not affected code is kept as is so as not to
    increase the diff even more;

  - Added missing comments to moved or changed structures and
    their attributes such as struct xrow_update,
    struct xrow_update_op_meta, struct xrow_update_op.

  - Struct xrow_update_field was significantly reworked. Now it is
    not just a couple of pointers at tuple's top level array. From
    now it stores type of the updated field, range of its source
    data in the original tuple, and a subtree of other update
    fields applied to the original data.

  - Added missing comments to some functions which I moved and
    decided worth commenting alongside, such as
    xrow_update_op_adjust_field_no(), xrow_update_alloc().

  - Functions xrow_update_op_do_f, xrow_update_op_read_arg_f,
    xrow_update_op_store_f are separated from struct xrow_update,
    so as they could be called on any updated field in the tree.
    From this moment they are methods of struct xrow_update_op.
    They take an op as a first argument (like 'this' in C++), and
    are applied to a given struct xrow_update_field.

Another notable, but not separable, change is a new naming schema
for the methods of struct xrow_update_field and struct
xrow_update_op. This is motivated by the fact that struct
xrow_update_field now has a type, and might be not a terminal.

There are now two groups of functions. Generic functions working
with struct xrow_update_field of any type:

    xrow_update_field_sizeof
    xrow_update_field_store
    xrow_update_op_do_field_<operation>

And typed functions:

    xrow_update_<type>_sizeof
    xrow_update_<type>_store
    xrow_update_op_do_<type>_<operation>

Where
    operation = insert/delete/set/arith ...
         type = array/map/bar/scalar ...

Common functions are used when type of a field to update is not
known in advance. For example, when an operation is applied to one
of fields of an array, it is not known what a type this field has:
another array, scalar, not changed field, map, etc. Common
functions do nothing more than just a switch by field type to
choose a more specific function.

Typed functions work with a specific type. They may change the
given field (add a new array element, replace it with a new value,
...), or may forward an operation deeper in case they see that its
JSON path is not fully passed yet.

Part of #1261

That patch finishes transformation of tuple_update public API to
xrow_update.

Part of #1261

Tuple_update is a too general name for the updates implemented
in these files. Indeed, a tuple can be updated from Lua, from
SQL, from update microlanguage. Xrow_update is a more specific
name, which is already widely used in tuple_update.c.

Part of #1261

When invalid command is passed we should send an error message to a
client. Instead a nil dereference occurs that causes abnormal exit of a
console.

This is the regression from 96dbc49d
('box/console: Refactor command handling').

Reported-by: Mergen Imeev <imeevma@tarantool.org>
Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>
Reviewed-by: Alexander Turenko <alexander.turenko@tarantool.org>

Added build + test jobs in GitLab-CI and build + test + deploy jobs on
Travis-CI for CentOS 8.

Updated testing dependencies in the RPM spec to follow the new Python 2
package naming scheme that was introduced in CentOS 8: it uses
'python2-' prefix rather then 'python-'.

CentOS 8 does not provide python2-gevent and python2-greenlet packages,
so they were pushed to https://packagecloud.io/packpack/backports
repository. This repository is enabled in our build image
(packpack/packpack:el-8) by default. Those dependencies are build-time,
so nothing was changed for a user. The source RPM packages were gathered
from https://cbs.centos.org

.

Disabled app-tap/pwd.test.lua on CentOS 8 due to systemd-nss issue,
which was not worked around properly. Filed #4592 to resolved it in the
future.

Eliminated libunwind runtime dependency (and libunwind-devel build
dependency) on CentOS 8, because the base system does not provide it.
fiber.info() backtraces and printing of a backtrace after a crash will
not be available on this system. Hopefully we'll fix it in the future,
filed #4611 on this.

Closes #4543

Reviewed-by: Alexander Turenko <alexander.turenko@tarantool.org>
Reviewed-by: Igor Munkin <imun@tarantool.org>

Type was displayed in error messages, was returned in
meta headers, and a type string is a result of
typeof() SQL function.

Typeof() always returns lower case type string; meta
contained upper case type; error messages contained
both.

It was necessary to choose one case for everything,
and the lower one was chosen. It allows not to break
typeof() function which actually might be used by
someone.

There was a bug that netbox at any schema update called
on_connect() triggers. This was due to overcomplicated logic of
handling of changes in the netbox state machine. On_connect() was
fired each time the machine entered 'active' state, even if its
previous states were 'active' and then 'fetch_schema'. The latter
state can be entered many times without reconnects.

Another bug was about on_disconnect() - it could be fired even if
the connection never entered active state. For example, if its
first 'fetch_schema' has failed.

Now there is an explicit flag showing the machine connect state.
The triggers are fired only when it is changed, on 'active' and on
any error states. Intermediate states (fetch_schema, auth) do not
matter anymore.

Thanks @mtrempoltsev for the initial investigation and a draft
fix.

Closes #4593

This function is supposed to return NULL on an error.
For exceptions there is user_find_by_name_xc.

Some guest user privileges were not revoked in the
end.

Explicit cast uses uppercase, and the patch makes the
implicit cast the same.

Upper case is the standard according to SQL standard
2011, cast specification 6.13, general rules 11.e:

    General rules

    11) If TD is variable-length character string or
        large object character string, then let MLTD
        be the maximum length in characters of TD.

        e) If SD (source type) is boolean, then Case:

            i)   If SV is True and MLTD is not less
                 than 4, then TV is 'TRUE'.

            ii)  If SV is False and MLTD is not less
                 than 5, then TV is 'FALSE'.

            iii) Otherwise, an exception condition is
                 raised: data exception — invalid
                 character value for cast.

Part of #4462

Before the patch LENGTH didn't take boolean argument
into account. Now it does and treats like any other
non-string argument - stringify and calculate length.

It is worth mentioning, that in future LENGTH will
discard any non-string argument, see #3929.

Part of #4462

Argparse module stores unspecified parameter values as boolean
true. It led to a problem, that a command line '--value' with
'value' defined as a number or a string, showed a strange error
message:

    Expected number/string, got "true"

Even though a user didn't pass any value. Now it shows 'nothing'
instead of '"true"'. That is clearer.

Follow up #4076

There was a complaint that tarantoolctl --show-system option is
very hard to use. It incorrectly parsed passed values, and
provided strange errors.

    tarantoolctl cat --show-system true
    Bad input for parameter "show-system". Expected boolean, got "true"

    tarantoolctl cat --show-system 1
    Bad input for parameter "show-system". Expected boolean, got "1"

    tarantoolctl cat --show-system=true
    Bad input for parameter "show-system". Expected boolean, got "true"

First of all, appeared that the complaining people didn't read
documentation in 'tarantoolctl --help'. It explicitly says, that
'--show-system' should go after a file name, and does not have a value.

Secondly, even having taken the documentation into account, the
errors indeed look ridiculous. 'Expected boolean, got "true"'
looks especially weird.

The problem appeared to be with argparse module, how it parses
boolean parameters, and how stores parameter values not specified
in a command line.

All parameters were parsed into a dictionary: parameter name ->
value. If a name is alone (no value), then it is boolean true.
Otherwise it was always a string value. An attempt to specify
an explicit parameter value 'true' led to storing string 'true'
in that dictionary.

Consequential check for boolean parameters was trivial:
type(value) == 'boolean', which was obviously wrong, and didn't
pass for 'true' string, but passed for an empty value.

Closes #4076

Credentials is a cache of user universal privileges. And that
cache can become outdated in case user privs were changed after
creation of the cache.

The patch makes user update all its credentials caches with new
privileges, via a list of all creds.

That solves a couple of real life problems:

- If a user managed to connect after box.cfg started listening
port, but before access was granted, then he needed a reconnect;

- Even if access was granted, a user may connect after box.cfg
listen, but before access *is recovered* from _priv space. It
was not possible to fix without a reconnect. And this problem
affected replication.

Closes #2763
Part of #4535
Part of #4536

@TarantoolBot document
Title: User privileges update affects existing sessions and objects
Previously if user privileges were updated (via
`box.schema.user.grant/revoke`), it was not reflected in already
existing sessions and objects like functions. Now it is.

For example:
```
        box.cfg{listen = 3313}
        box.schema.user.create('test_user', {password = '1'})
        function test1() return 'success' end

        c = require('net.box').connect(box.cfg.listen, {
                user = 'test_user', password = '1'
        })
        -- Error, no access for this connection.
        c:call('test1')

        box.schema.user.grant('test_user', 'execute', 'universe')
        -- Now works, even though access was granted after
        -- connection.
        c:call('test1')
```

A similar thing happens now with `box.session.su` and functions
created via `box.schema.func.create` with `setuid` flag.

In other words, now user privileges update is reflected
everywhere immediately.

(cherry picked from commit 06dbcec597f14fae6b3a7fa2361f2ac513099662)

Trigger function returning type is changed from void to int and
any non-zero value means the trigger was processed with an error.
A trigger can still raise an error - there is no more refactoring
except obvious `diag_raise();' --> `return -1;' replacement.

Prerequisites: #4247

Currently, the test shows all the data contained in the spaces
_func, _user and _space. This was added 6 years ago and is no
longer needed. In addition, this is inconvenient, as some data
changes every time bootstrap.snap is created. Due to this, we must
update the test result file every time we generate bootstrap.snap.
This patch removes the display of this data from the test.

Before the patch there was a race in replication
password configuration. It was possible that a replica
connects to a master with a custom password before
that password is actually set. The replica treated the
error as critical and exited.

But in fact it is not critical. Replica even can
withstand absence of a user and keeps reconnecting.
Wrong password situation arises from the same problem
of non atomic configuration and is fixed the same -
keep reconnect attempts if the password was wrong.

Closes #4550

Closes #4519

@TarantoolBot document
Title: key_def.new() accept both 'field' and 'fieldno'

Before the patch key_def.new() took an index part
array as it is returned in <index_object>.parts: each
part should include 'type', 'fieldno', and what else
.parts element contains.

But it was not possible to create a key_def from an
index definition - the array passed to
<space_object>.create_index() 'parts' argument. Because
key_def.new() didn't recognize 'field' option. That
might be useful, when a key_def is needed on a remote
client, where a space object and its indexes do not
exist. And it would be strange to force a user to
create them just so as he would be able to access

    <net_box connection>.space.<space_name>.
        index.<index_name>.parts

As well as it would be crutchy to make a user manually
replace 'field' with 'fieldno' in its index definition
just to create a key_def.

Additionally, an ability to pass an index definition
to a key_def constructor makes the API more symmetric.

Note, that it still is not 100% symmetric, because a
user can't pass field names to the key_def
constructor. A space is needed for that anyway.

Before the patch the nil UUID was ignored and a new random one
was generated. This was because internally box treats nil UUID
as its absence.

Now a user will see an explicit message that nil UUID is a
reserved value.

Closes #4282