- Use 'static inline' instead of 'inline'
 - Move GROW_INCREMENT enum out of LIGHT(record) struct

So as not to confuse the user, let's start reader threads when recovery
completes only if there are Vinyl indexes. Otherwise, start them on the
first Vinyl index creation.

See #2664

Vinyl starts its worker threads on checkpoint, even if there is no Vinyl
spaces to dump out there. This is a mess. Let's wake up the scheduler
only if we actually have something to dump. BTW, this patch removes that
annoying 'vinyl checkpoint done' message in case nothing was dumped.

See #2664

Introduce the following functions:

  vy_scheduler_begin_checkpoint()
  vy_scheduler_wait_checkpoint()
  vy_scheduler_end_checkpoint()

and make the Vinyl API methods use them instead of playing with the
scheduler directly.

Rationale:
 - It will help move the scheduler to a separate file in future.
 - We want to avoid waking the scheduler on checkpoint if Vinyl is not
   used, because the scheduler starts worker threads on the first
   wakeup. At the same time, we have to call vy_log_rotate() on
   checkpoint no matter if Vinyl is used or not (it needs to update its
   internal housekeeping). Factoring out the scheduler functions will
   allow to achieve this goal neatly.

There's no point in maintaining the LSN of the last checkpoint in Vinyl,
we have checkpoint_last() for retrieving it. Make this helper a little
bit more friendly by allowing to omit vclock argument and use it
wherever appropriate.

Here's why I need to do that now. I'm planning to omit calling scheduler
checkpoint functions altogether if there's nothing to dump (so as not to
start worker threads and print 'vinyl checkpoint done' message in case
Vinyl is not used), but one of those functions updates last_checkpoint,
which other pieces of code (gc primarily) relies on. I could factor
last_checkpoint out of vy_scheduler, but why if I can simply use
checkpoint_last() instead.

Currently, box_set_checkpoint_count() not only changes the corresponding
configuration option, but also invokes garbage collection. The problem
is it is called in the beginning of box_cfg(), before recovering engine
data. At that time engines are not ready to invoke garbage collection
(e.g. Vinyl hasn't scanned vylog directory yet). Besides, the server
might be started in the hot standby mode, in which case calling garbage
collection is obviously wrong.

To fix this, we could somehow detect if box_set_checkpoint_count() is
called from box_cfg() and avoid invoking the garbage collector if so,
but I think it isn't worth it. Instead let's just not invoke garbage
collection when this option changes. It's not a big deal - old files
will be removed by the next snapshot anyway.

It's inappropriate to fail startup if the vinyl data directory does not
exist, because vinyl might not even be in use. Move the check to the
time when a vinyl index is opened.

See #2664

If an index is dropped while a dump/compaction task is in progress, the
task will be aborted, but VY_LOG_DROP_RUN record won't be written to the
metadata log so the run will remain incomplete (VY_LOG_PREPARE_RUN) and
pin the dropped index in the log until the server is restarted and all
incomplete runs are purged. This results in occasional failures of the
vinyl/gc test which checks that vylog files are removed after all vinyl
spaces are dropped. There's actually no reason not to log deletion of an
incomplete run that belongs to a dropped index so let's do this to fix
this problem.

Closes #2486

So that BPS tree can be redefined in files that include memtx_tree.h

Fixes #2761

Apple's version of libcurl is outdated and buggy.
Use libcurl from homebrew by default.

Closes #2772

queue.start() is not needed anymore.

Follow up the previous commit.

The cache iterator restoration function is muddy and suboptimal:

 - If last_stmt equals NULL, it doesn't attempt to change the iterator
   position although it should restart the iterator. Moreover, it
   doesn't even check EQ condition in this case, i.e. it may return a
   statement that doesn't satisfy the search criteria.

 - It may restore iterator position to last_stmt, but the restoration
   semantics mandates that the returned statement must be strictly
   greater than last_stmt if the iterator was restored.

 - If the iterator was invalidated, it first positions the iterator to
   the last known position (curr_stmt) and then steps back to the
   statement closest to last_stmt although it could jump to last_stmt
   immediately.

Let's rewrite it.

vy_merge_iterator_squash_upserts() returns an UPSERT, which is then
converted to a REPLACE by vy_read_iterator_next(). Fold the latter into
the former to make the workflow easier to follow. Also, remove useless
suppress_error argument.

Not only restore() callback can restore the iterator position after the
source changed - next_key() and next_lsn() can do that too, which brings
additional complexity. Do we really need it? Not really:

 - We can avoid handling restoration in next_key() by simply calling
   restore() before it. This is cheap as restore() is a no-op if the
   source was not changed.

 - Regarding next_lsn() - when it comes to calling it, we iterate over
   sources from the most recent one to the oldest, which means that the
   only sources out there that may yield, on-disk runs, are guaranteed
   to be iterated last, after all mutable sources (txw, cache, mems), so
   none of the mutable sources needs to be restored in this case.

Simplify the read iterator as per above.

Currently, ->next_lsn() may be (and actually is) called right after
opening the iterator, which means that apart from advancing the iterator
to the next version of the same statement it has to handle the case of
starting iteration. This complicates the callback semantics so let's
require ->next_key() to be called first.

Closes #2774

We generally do not allow to drop objects that have dependents.

In contrast to PostgreSQL, this method doesn't make sense in Tarantool,
because we don't have sessions.

This patch implements a new object type, persistent sequences. Sequences
are created with function box.schema.sequence.create(name, options).
Options include min/max, start value, increment, cache size, just like
in Postgresql, although 'cache' is ignored for now. All sequences can be
accessed via box.sequence.<name>, similarly to spaces. To generate a
sequence value, use seq:next() method. To retrieve the last generated
value, use seq:get(). A sequence value can also be reset to the start
value or to any other value using seq:reset() and seq:set() methods.

Needed for #389

Currently, it is done by space_cache_replace/delete which violates
incapsulation. Let's introduce a trigger that is fired after a change in
a space definition is committed and use it to propagate changes to Lua.
Patch by @kostja.

Implement a helper function to swap space triggers.
Patch by @locker.

Speed up zstd as much as possible.

Update zstd submodule to v1.3.1 version.

Needed to implement SMTP client on the top of libcurl.

See #2704

We can't call initial() from 1.7.6+ because of the space format checks.
Port initial_1_7_5() from 1.8 and use it to boostrap.

Coverity Scan has found a possible bug in a CSV parser - dereferencing
of a null pointer.
But it is not possible. Lets explain:

According to Coverity, the null dereference is in CSV_IN_QUOTES
processing in csv_parse_impl.
But bufp (nullified pointer) can be set to null only in two cases:
CSV_NEWFIELD and CSV_OUT_OF_QUOTES.

In a case of CSV_NEWFIELD the parser sets the state to CSV_LEADING_SPACES and
on a next iteration it sets bufp to not null.

In a case of CSV_OUT_OF_QUOTES the parser can nullify bufp only
if is_end_of_line was true at the end of 'out_of_quotes' processing.
But it means, that is_end_of_line was true at the beginning of
processing too, because CSV_OUT_OF_QUOTES does not modify it.
And at the beggining, if is_end_of_line was true, then the state
is set to CSV_LEADING_SPACES. And on a next iteration bufp are set
to not null.

Closes #2692

Closes #2754

Needed for #2754

sc_space_new return value always is unused. Make sc_space_new
return void.

Needed for #2754

Version is needed in #2754 to ignore space format,
if schema version < 1.7.6.

Needed for #2754

Needed to get format fields during alter in #2754