Add test cases for transaction control statements
invoked in a trigger.

ABI change: box_txn_rollback() now may fail with error,
if called from a nested stateent.

Touch txn a little bit to support multistatement transactions in
autocommit mode.

This patch allow to perform DML requests from space:on_replace()
triggers. Trigerred requests are processed as a part of the original
transaction and entire transaction is rolled back in case of error.
Please note that there is still no support for nested transactions,
therefore a call to box.begin() from on_replace() trigger is not
allowed.

Recursive invocation is not forbidden, but limited to the maximal
depth of 4 to avoid stack overflow.

Proudly fixes #587

Closes #1378

Fixes #1962

Needed for #1692

Rawhide and Sid are bleeding edge technlogies.

Remove hardcoded paths.

Now as we have vy_scheduler->mem_min_lsn, we can check if checkpoint has
finished by comparing it to ->checkpoint_lsn, so ->checkpoint_pending
and ->dirty_mem_count fields are not needed and can be removed.

Also, remove vy_scheduler->checkpoint_failed field, because ->throttled
can be used instead.

Required to implement new chunked allocator for tuples.

Closes #1950

Required to implement new chunked allocator for tuples.

Closes #1948

Add rlist_splice.

If there's a dropped index referenced by a pending task on scheduler
stop, vy_scheduler_stop() will crash while trying to delete it:

  vy_scheduler_stop
   vy_task_delete
    vy_index_unref
     vy_index_delete
      vy_range_tree_free_cb
       vy_scheduler_remove_range

because there will be a range which is not on the dump/compact heap.

Fix that by completing all processed tasks and aborting (i.e. setting
status to -1 and calling ->complete) all pending tasks when stopping the
scheduler.

Always write .run files even if there are no records.

When we finally start storing vinyl metadata in a memtx table, we will
create an entry for a run as soon as we add the in-memory index which is
going to become the run after it's dumped. On recovery, we won't replay
xlog rows which add statements to in-memory indexes that have already
been dumped. We are planning to detect this by checking if the in-memory
index has a file on disk. To be able to do this in case all entries in a
range cancel each other on dump, we need to create files for empty runs.

Note, although we could delete empty runs from memory right away, we'll
need to have all runs when we switch to storing vinyl metadata in memtx,
including empty ones, on the list so that we can delete records from the
memtx table when they get compacted. In order not to break read/write
iterator assumptions, just skip empty runs when opening an iterator.

coredump() feature is broken, see #1886

Blobs and minified JS are prohibited by distributions.
Fixes packages on Debian Sid.

Closes #1852

Time calculation is too accurate.

Fixes #1946

Fixes #1896

Extend the conflict verifing test
Fix gh-1913

We collect garbage by deleting all files in the index directory that
aren't referenced by any range (see vy_index_gc). We assume a file to be
referenced, if there's a range in the index->tree whose id matches the
file name. Since shadow ranges (used for compaction) are expelled from
the tree, their files might be removed as well, which would result in
data loss if compaction was aborted. Fix that by adding shadow ranges to
the the temporary map used for filtering out active ranges.

* Ignore ER_ACCESS_DENIED in applier
* Replace warning_said flag with the errcode of the last error to
  suppress log messages properly
* Supress noise "connected to xxx:yyy" messages which floods the log

A test case was copy-pasted due to limitation of test-run.
See https://github.com/tarantool/test-run/issues/40.

Fixes #1589

During the first call box.cfg() now waits for an eternity for all
cluster members to start up. All subsequntial box.cfg() calls will
timeout after 10 seconds if one or more hosts are unreachable.

Fixes #1511

Extract vy_stmt_encode and vy_stmt_decode as separate functions to
read/write xlog files from vinyl.

Issue #1720

* Add TRASH(xlog) to xlog_close()
* Fix errno in case of error in xdir_create_xlog()
* Swap arguments of xdir_create_xlog()

Issue #1720