Add missing -isysroot flag for LuaJIT on macOS.

Fixes #2091

Fixes #2189

We can't write to an xlog from a coeio task, because coeio uses a pool
of workers while xlog can only be used from a single thread due to obuf.
So we use a hack: we write to the metadata xlog in the tx thread, but
flush the xlogfrom a coeio task so as not to block tx. To get rid of it,
let's write the metadata log on behalf of the wal thread.

We are planning to share the metadata log among all engines and the
tarantool core, so rename vy_log appropriately.

For the sake of encapsulation, we don't remove run files directly from
vy_log.c - instead we pass a gc callback on log creation. Once we make
the metadata log system-wide, so that it could be used for storing info
about snapshots and xlog files, we won't be able to do that. So in this
patch we remove vy_log_gc_cb callback and move run file deletion to
vy_log.c.

When a vinyl index is dropped, there still might be dump/compaction
tasks running for it. Upon completion a task will log run/range creation
or deletion, which will cause recovery to fail if the index is dropped.
To avoid this, let's avoid logging metadata after the index was dropped.

Follow up #2005

Combine rmeans for VY_STAT_UPSERT_OPTIMIZED, VY_STAT_UPSERT_OPTIMIZED_TX and
VY_STAT_UPSERT_CHAINS_OPTIMIZED. We don't need such detailed statistics.

Count for squashed upserts, in-transaction replaced upserts and all
applied upsert operations.

* do not allow "deleting" id of self from _cluster table
* assume struct replica for self is always present, even
  when replication is off
* update cluster.test.py, remove tests for removal of self
  instance id
* use struct replica self->id instead of global instance id
  wherever necessary
* read only mode is now defined by is_ro alone, since
  the local instance always has an assigned id

This cherry-picks some changes from branch remove-tx-vclock,
and follows up on the removal of recovery->replica_id member.

Ensure we always have at least "self" in the replication cluster
(replica set), even if replication is off.

The local instance can never be removed from the replica set.

This simplifies further reasoning about the code.

Add a new instance_id variable to identify position of the local
instance lsn in vclock vector.

See #1927

* Update and re-wrap example.lua
* Update /etc/default/tarantool
* Add support for old names in /etc/default/tarantool

See #1927

Fix type for old parameter name for config update.

Plain vinyl options. Rename memtx memory options and remove GB scale.
Rename replication source, log* and no_panic_* options.

* Rename functions in tuple_compare.cc
* Add NOINLINE function attribute to util.h

See #2122

Closes #2112

The original fix makes sure that applier does not overwrite stop
vclock with current vclock read from the end of final join stream.

This patch builds upon the original fix and never uses the
vclock from the end of the final join stream, since it's not
used at the moment.

After final join replica should subscription from the end of recieved
final join stream but not current master vclock value after the final join
was sent.

Fix the 'abort' callback for compaction and split tasks as well.

In case of shutdown vy_scheduler_stop_workers() calls task->abort
with empty task->diag.

Closes #1908

We need to explicitly store the format of a space
and count of indexes in the index structure, since
the space can be dropped before deletion of an index
and index->space pointer will be invalid.

We have been skipping the DELETE statements, which
were first DELETEs in an index, but such a case
is possible only once at index life time.
Also this optimization complicates the code and
intruduces additional overhead on each
vy_ranse_set_delete.

Another problem of this optimization is that we
in advance can't know will or will not the
DELETE statement be inserted, before calling
vy_range_set_delete.

To simplify the code and prepare to common vy_mem,
this optimization is removed.

If all worker threads are busy doing compaction when we run out of
quota, ongoing transactions will hang until one of the threads has
finished, which may take quite a while. To avoid unpredictably long
stalls, always keep one worker thread reserved for dumps.

Closes #2150

This ensures tuple cache is used in standalone select{} and
get{} requests.

Remove RO tx completely. Fix a bug in the cache iterator.

Needed by memtx in order not to violate incapsulation.

Remove the check for actual page count exceeding the one written
in page meta: the check was incorrect.
It took into account a possibl extra row in xlog transaction,
but didn't check if the log has other transaction.

The check for extra transactions is not trivial.

The check for exact match itself is not necessary for correctness,
so simply remove until xlog API lends itself more easily to such
check.

See also gh-2157

1. Do not ignore "too few pages" error during run recover.
2. Differentiate EOF error and parse error during run recover.
3. Execute region cleanup after dumping a statement to a run.

fiber_join checks fiber_is_canceled flag and do not raise for this case.
So, if fiber was already dead we should not set canceled flag for the fiber.

Improve vinyl task logging.

Implement cpipe_destroy() and cbus_endpoint_destroy(). cpipe_destroy()
send a poison message to endpoint and flushes all queued. Each cbus_endpoint
counts the number of connected pipes and doesn't destroy if there are
connected pipes or unfetched messages.

cbus_endpoint_create() returns an error instead of panic in case of a
duplicate endpoint. A new endpoint can be created with same name while older
endpoint still destroing.

Fiber pool destroys endpoint, resets idle timeout to 0 and gathers all
workers.

Null-merge changes from 1.5.