Fix bugs in spawner child handling, performed at shutdown:
- shutdown children more portably
- remove races with lost SIGCHLD
- do not self-terminate when doing kill(0, signal)
(sending signal to entire process group).

Clean up custom_proc_title handling and fiber names.

Allow fiber_io_* on an fd which is ! fiber->fd,
and remove inbox2sock/sock2inbox in replication.

Fix a merge error (remove debug print).

Remove one memory leak in recovery_follow_remote.

Rename hot_standby back to local_hot_standby. Let's
not break compatibility without a strong need.

Merge replication_source_ipaddr and replication_source_port
to a single variable, replication_source, to be able to
provide a list of sources in future.

Update comments in the configuration file.
Remove custom proc title for replication, use custom_proc_title of
the main process.
Renames.

Conflicts:
	cfg/tarantool_box_cfg.c
	cfg/tarantool_box_cfg.cfg
	cfg/tarantool_box_cfg.h
	cfg/tarantool_feeder_cfg.c
	cfg/tarantool_feeder_cfg.h
	core/exception.m
	core/fiber.m
	core/log_io.m
	core/log_io_remote.m
	include/fiber.h
	mod/box/box.h
	mod/box/box.m
	mod/box/box_cfg.cfg_tmpl
	mod/box/memcached.m
	mod/box/memcached.rl
	mod/feeder/feeder.m
	test/box_replication/common.result
	test/box_replication/suite.ini
	test/lib/admin_connection.py
	test/lib/box_connection.py
	test/lib/server.py
	test/lib/sql_ast.py
	test/lib/tarantool_box_server.py
	test/lib/tarantool_connection.py
	test/lib/tarantool_feeder_server.py
	test/lib/tarantool_server.py

Blueprint 'Make memcached protocol an alternative, not a replacement':

Blueprint was implemented. Memcached now can work
in presence of RELOAD CONFIGURATION statement, even if it
changes a replica to master or a master to replica.

'memcached' configuration file option was removed.
Instead, two new optinos were added:
- memcached_port - optional, default 0, if set, specifies
which port speaks memcached protocol.
- memcached_expire, boolean, which defines whether or not
we run expire on memcached namespace. If not set,
we simply support the protocol, but ignore TTL and never
expire the keys.

Original memcached test was ported to tarantool testing environment.
cas test was disabled, because tarantool has a non-critical bug (in
concurrent environment cas operation can return 'SERVER_ERROR' instead of
'EXISTS')

Handle close replication socket on oposite side was added.
box replication test was updated.

Missed inotify even in the replicaor handler process was foxed.
replicaor_init was splited on two function:
  - replicaor_prefork -- create replicaor spawner process;
  - replicaor_init -- initialize replicaor fibers in the main tarantool
	                  process.

Test and test configuration files in box_replication soute was
clenupped and updated.

We modify memcached code quite often,
and asking ragel to regenerate .m from .rl file
every time .rl file changed was tiresome both
for revision control and for code reviews.

Leave in memcached-grammar.rl only memcached_dispsatch(),
which is generated by ragel, and everything else
move to memcached.m. This should lessen cases
when we have to re-generate memcached-grammar.m.

This patch doesn't produce any code changes,
only moves the code (and cleans up include
files at start of memcached.m).

Remove compilation warnings about unused variables.

Make sure Release and RelWithDebugInfo do
not include -Werror - we don't know compiler version
at the client and shouldn't plant time bombs for him/her.

box replication test was updated:
 - common test was splited on three test:
    * hot_standby test -- check replication from hot_standby server
        * swap -- this test iteratively swap services roles (master to
                  replica, replica to master)
        * consistent -- this is test check replica data consistent
 - unused configuration files was deleted.
 - get_param method to tarantool_to_slave

Replicator was updated:
 - bind replicator poer strategy was changed.
 - terminate child routine for spawner process during exit was added.
 - handle EPIPE error was added to client handler process.
 - remote_hot_standby was deleted.
 - local_hot_standby was renamed to hot_standby.
 - sender_send_sock function was fixed.
 - I/O event processing in replicator fibers was updated.

Fiber:
 - fiber_io_* functions now public.

Reorder and simplify calls to box_enter_master_or_replica_mode()
in mod_relaod_config().

Move recovery_stop_remote() to log_io_remote.c, where it belongs.
Remove now unnecessary recovery_start_remote(),
use recovery_follow_remote() (same thing) instead.

mode

Reorganize the code responsible for entering/leaving master mode.

Remove remote_recovery_restart, add remote_recovery_{stop/stop},
make sure we know what to do - start or stop recovery.

Move 'remote_recovery' to struct recovery_state.

Remove an extra parameter to recover_follow_remote,
it's always the same currently.

Fix a coding error: supplying 0 watermark to kh_init(fiber_registry)
lead to almost constant kh_resize (on every kh_put after a kh_del).

Fix the memory leak which would occur when
allocating memory for the "status" string,
output by 'show info'. The allocation was done
on each RELOAD CONFIGURATION, and
old memory was never freed.

Fix: use static memory instead of palloc.

PEP8 coding-style for Python code was added. Python code ndentation
was changed to PEP8 format.

Make sure we do not have pending updates prior to turning
on remote recovery.

Get rid of the boolean variable 'box_updates_allowed'
and pass a function pointer to iproto_interact
instead (one less 'if' on the way of the most common
execution flow).

Remove diagnostics.[hm] since we should use
Objective C exceptions going forward.

A few more cleanups related to v03/v04 snapshot/wal
format.

Basic cleanup that removew old/unused code.

Send error messages to the client.

Before this patch, Tarantool would only send an error
code to the client, but record all the details in its error log.

This patch extends the client/server protocol and adds
error text to it. This text can now contain not only
a staple message that corresponds to the error, but
problem-related additional information, such as namespace id,
field number, etc.

All errors still can be logged in the error log, but
not all of them need to be. For errors, which are only
sent to the client, this patch uses ClientError exception
class. Errors which also need to be logged in the error log
must use LoggedError exception class.

To choose between the two, the following rule of thumb is applied:
- client application errors, which, generally, can be non-errors at all,
  such as access to a non-existing namespace, index or field, or
  access to a locked tuple are not logged
- execution and protocol errors, such as an error writing to disk,
an message that doesn't follow the protocol format, is logged.

For the most common error, ER_TUPLE_IS_RO, we do not use
string template to reduce CPU cycles necessary to produce this error
(on my machine one can raise ~200 000 exceptions per second, which is not
too much).

An addition to the above, this patch renames error codes, to
make them shorter (ERR_CODE_* -> ER_*) and rewrites the iproto.m
client connection loop to make it a trifle more efficient.

This patch also fixes a couple of bugs:
- in exception.m, [tnt_Exception alloc]
when we reused the old exception class, we did
not reset its 'isa' pointer, which could lead to a crash
when a message with a different signature was sent to
it.

- in box_dispatch (renamed from box_dispach),
when logging for slow commands was done, txn->op could
already be overwritten by txn_commit()/txn_abort().

A comment on the naming scheme: prefixing every Tarantool
piece of method/data structure with tnt_ would lead
to very bloated code. So in this patch
I decided to reduce the use this
prefix to the client library code and some very
generic names only.

When stopping the server under gdb, send SIGTERM
to the server, not to the terminal in which
gdb is started. This gives you the option
to actually examine gdb output before the test
has ended.

When the server crashes, and we're reading
from it in a client connection, detect EOF
and abort reading (the old code was buggy
and would busyloop indefinitely in case of
a crash).

Remove an unnecessary fork() in
start().

When ctrl-c is pressed on the terminal, a signal (SIGINT)
is sent to all members of the foreground process group.

Since both wal_writer and logger processes' default action
on SIGINT is to terminate, they could terminate before parent.

This led to rather nasty results: in the circumstances, when
the parent somehow ignored SIGINT (e.g. under GDB), or handled
it after its children (a race condition), children would
disappear before parent, then parent would detect that
wal_writer is gone, try to panic/write to the log, which
is also gone, and on top of SIGINT get a SIGPIPE.

Move spawned children to their own process groups
to make sure they don't belong to the main process'
process group, and don't receive signals from the
controlling tty. It's okay in the interactive mode,
since these children are going to die anyway when the
parent dies, after getting a SIGPIPE.

Do not try to start another server on a port if it's already in use.
This can lead to very confusing results when running
tests that start and stop the server in parallel.

Remove an extra space in '-e ' option to term,
not all terminals (rxvt, for example) can handle the extra
space in an option name.

Update *.cfg files to newer confetti and comments
in the templates.

Rename exceptions.[hm] to exception.[hm]. Add copyright
information.

There were extra quotes in errcode.c, i.e.
quotes were part of the error message.
Remove them.

Sweep out the first set of coding bugs in the test suite:
- set slab_alloc_arena to 0.05 in all box_replication configurations,
since otherwise this suite tries to allocate too much memory
and hangs under a VM.
- update some texts that make it into the log (spelling)
- add logging to all configurations
- simplify wait_sync and rename it to wait_lsn
- there was a bug in wait_sync that time.sleep() was
not called in the body of the loop due to \t, and
the loop was a busy-wait one.