Add helpers to decode numbers and use them throughout the code.

tuple_update/upsert was inlined in memtx_space and
box_tuple_update/upsert.

Fixes #1778

Needed to ensure that update_optimize will work
after implementation of common memory level
for txv and vy_mems.

Needed for #1908

Exactly the same piece of code (including comment) used by both dump and
compact. Factor it out into a separate function.

All the three functions that need a write iterator, dump, compact, and
split, reuse the same code for initialization of the write iterator.
Factor it out into a separate function.

* Use SNPRINT() macro from util.h
* Prettify formatting with missing schema (e.g. "user@localhost")
* Add Lua test case

Fixes #1924

1024MB is a lot for small devices and virtual machines.
256 MB is more than sufficient to play with a database,
and will make life easier when trying the database for
the first time.

Closes #1692

Change order of traversal in merge iterator to direct,
getting ready for avoiding reading from disk in case of cache hit

Required to check if a range should be coalesced with its neighbors.

 - Do not unaccount mem from the index and unlink it from the scheduler
   in vy_range_delete(), because it violates encapsulation. Let's just
   delete it there and leave accounting to specialized helpers, as we do
   in case of runs.
 - Delete vy_range_delete_mem() helper, because it has nothing to do
   with range - it unaccounts a mem from the *index* and removes it from
   the *scheduler*, and then frees it.
 - Make vy_index_{acct,unacct}_range() helpers handle mems as well, as
   those are a part of each range.
 - Drop vy_index_acct_range_dumped() helper. Use unacct/acct pair
   instead.

Hopefully, ths patch makes the code easier to follow.

Besides vy_run_total(), which returns the size of run data file on disk,
there's also vy_run_size() which returns the size of run and page info
structs in memory. The latter doesn't make much sense, so let's drop it
and rename vy_run_total() to vy_run_size().

 - Generate random keys to prevent compression.
 - Expect 2 ranges, not 4, because we do not split a range unless
   it was compacted. We could make the test generate 4 ranges by
   issuing snapshot more often, but that would increase execution
   time, so leave it as is.

Follow up a49673c9

* update the default level_size_ratio
* rename level_size_ratio to run_size_ratio
* rename max_runs_per_level to run_count_per_level

To follow the LSM tree design, group runs of the same range into levels.
The division is ephemeral, it is recomputed after each dump and not
stored in vy_range struct. The size of the first level is set to the
largest run dumped since the last compaction. The size of each
subsequent level is level_size_ratio times larger than previous one.
When the number of runs at a level exceeds max_runs_per_level, all runs
of the maximal level that exceeds the limitation as well as all runs and
mems of preceding levels are compacted. max_runs_per_level and
level_size_ratio are set in index config. By default, they equal to 2
and 10, respectively.

Closes #1821

Fix naming convention.

Follow up the previous patch.

Add a workaround for suboptimal JIT performance on this test.

See #2039

Currently, we create a run file and add a new run to the list even if
it's empty and therefore there's no point in keeping it. Besides being
pointless, this feature also stands in the way of multilevel compaction,
where dumped runs are supposed to define the size of the first level of
the LSM tree. If a run is empty, the math behind the level size
computation does not work and should be special-cased, which look ugly.
That said, let's just drop empty runs right away.

Currently, range compaction task always compacts all runs and mems, and
so shares the code base with the split task. However, this is going to
change once multi-level compaction is introduced, so we need to separate
them.

Rename compact_list to split_list and fix the associated comments.
This is a preparation for separating range compaction from split,
which is necessary for multi-level compaction.

Fixes #1766

Also, let's start allocating IDs from 0, not from 1.

Note, for the latch object to be suitable for the vy_log use case, we
make it remember the fiber that locked it. While we are at it, let's
also fix some comments.

See #1488

Closes #1963

Also, log ERRINJ_VY_GC induced error.

 - Do not print range id and index lsn as they don't make much sense any
   more (don't point to a file), only print range boundaries and space
   and index id.
 - In case of split print the key by which it is done.
 - Remove debug messages for range creation/deletion, because with the
   split key being printed on compaction they are redundant.
 - Also, while we are at cleaning up log messages, let's use say_warn()
   instead of say_info() for notifying the user that the scheduler was
   throttled, because it is rather abnormal.

Refactor vy_task_dump_new() to make it more straightforward.
No functional changes intended.