* Remove dependecy on `struct xlog` from `struct xlog_cursor`
* Use file descriptor instead of FILE *
* Extract `struct xlog_meta`

If we managed to get the request length, we don't need to close the
connection on decode failure.

Note in test/box/net.box.test.lua: in order to make the server close the
connection, x_fatal() now has to send a request which doesn't start with
MP_UINT.

Closes #1654

Need for #1572

* many new comments
* refactoring vy_stmt methods
* remove dead code

See #1572

See #1572

vy_stmt_is_not_found was used for checking if vy_index_read
returned DELETE, but it is impossible - vy_index_read uses
vy_read_iterator_next that always returns REPLACE.

Error injections are not enabled in Release mode.

2Mb tuple is overkill for 1k page.

Fixes #1868

RB_CMP_TREE_ARG isn't necessary as we can extract key_def from
txv->index or range->index.

Closes #1754

Follow up previous commit

* Bug#842086: tarantool-dev: removal of tarantool-dev makes files disappear
  from tarantool-lts-dev
* Bug#842073: tarantool-common: unowned symlink after purge
* Lintian warning about debian/compact version

A regression caused by 7d12d66e

net.box returns true on :ping() even on ER_WRONG_SCHEMA_VERSION.
In other words, PING is not re-tried by net.box.

Fixes #1856

Evaluating arguments can be expensive. We should avoid it if we are not
going to emit a log message due to the log level.

Closes #1851

This:

  rmean_mean(s->rmean->stats[VY_STAT_TX_WRITE].value);

looks cumbersome. Let's rewrite the rmean api to make it look like this:

  rmean_mean(s->rmean, VY_STAT_TX_WRITE);

Also, introduce rmean_total() helper.

Follow up bae33669

 - Propagate dump/compact errors from workers to vy_scheduler->diag via
   vy_task->diag and log errors from the scheduler.
 - Abort vy_wait_checkpoint() from the scheduler if it fails to schedule
   a task or a task completes with error.
 - Throttle the scheduler whenever it encounters an error, so as not to
   burn CPU by rescheduling the same faulty task over and over again.
   The timeout grows exponentially with each successive failure, min 1
   second, max 60 seconds.
 - Abort vy_checkpoint() if the scheduler is throttled.
 - Rename ERRINJ_VY_RANGE_CREATE to ERRINJ_VY_RANGE_DUMP and abort any
   dump if it is set in order to make vinyl/errinj test predictable
   (we don't know when a range is actually created - it may happen
   before box.snapshot() - so the latter isn't bound to fail if
   ERRINJ_VY_RANGE_CREATE is set).

Closes #1857

* Set diag on mempool_alloc() failure
* Pass `struct mempool` instead of `struct vy_env` as an allocator context

Follow up #1767

We should probably create a separate arena for vinyl or reuse the memtx
arena. For now, allocate from the generic cache.

Closes #1767

Closes #1860

Closes #1848

Synchronize naming with the latest C++ standard.

Needed for #1848

This patch doesn't introduce any functional changes.

See #1585

Closes #1585

Closes #1131

* added 'ignore_crc' flag to xlog_cursor
* added tests for bad crc, bad version, bad format, good v12/v13 fformat reading

+ Review fixes from rtsisyk

Fixes #1854

Currently, an error is propagated from Engine::{begin,wait}Checkpoint()
to lbox_snapshot() in errno. This is ugly and inflexible, as we are only
allowed to throw SystemError from Engine - see how
MemtxEngine::waitCheckpoint converts diag to SystemError in order to
extract errno. We should use diag instead of errno all the way from
Engine to lbox.

For reporting about checkpoint-in-progress, use ClientError with a new
code, ER_SNAPSHOT_IN_PROGRESS, instead of EINPROGRESS errno, as it's
obviously not a SystemError.

Closes #1858

See #1572

xlog for each recovey after new xlog file and pread implementation.
Optimize fake eof (not closed xlog) cause for recovery.

xlog for each recovey after new xlog file and pread implementation.
Optimize fake eof (not closed xlog) cause for recovery.