* make fiber.test.lua stable
* make snapshot.test.lua stable
* return consistent error code (SystemError) from unrecoverable
  xlog errors

Fixes #2006

Fixes valgrind warnings.

- Don't break encapsulation of xlog
- Remove tuples cache from vy_page
- Use mp_store_u32()/mp_load_u32() to encode/decode row index
- Fix error handling

Add row index to vinyl page. Row index contains row offsets for the
page. We don't need to decode full page to read each tuple. Construct
vy_stmt from page on demand.

Issue #1954

The work done by upsert squash fiber should be proportional to the cost of
reading a page from disk.

Before dumping a range, we create a new mem so that it would accommodate
statements to be inserted while we are writing the rest of mems (see
vy_task_dump_new()). Actually, we can skip this step in case the newest
mem is empty. Besides being reasonable, this micro-optimization will
come in handy for the patch moving vinyl metadata to memtx.

Instead of surrounding every call to vy_scheduler_mem_dumped() with the
code handling quota update.

- Check "RUN" and "INDEX" meta types.
- Add error codes for vinyl xlog checks.
- Handle xlog meta version internaly.

Fixes #1952

 - Use vy_key_data() or vy_tuple_data() depending on
   the statement type.
 - Print operations for UPSERT statements.

* clean header includings
* rename some public vinyl functions
* merge VinylPrimaryIndex and VinylSecondaryIndex

See #1908

Self-explaining fix. Also, add an assertion to vy_quota_release() to
catch this kind of errors earlier.

Vinyl is going to have some peculiarities regarding replication,
so we should make the replication test suite try both engines.

Closes #1997

* update comments, minor tweaks

* merge lookup_full_tuple into vinyl.c
* remove VinylSecondaryIndex key_defs

See #1908

* minor style edits
* remove an unused function as we go

Split service start to separate functions bind and listen. Remove rebind
logic from evio_service and on_bind callback.

See #1908

This patch is a quick and dirty fix for the following commits:

  3bda8943 vinyl: use one fiber for optimizing upserts
  e132110b vinyl: avoid scheduling upsert optimizer after each upsert
  3940cba2 vinyl: do not check index/range version when optimizing upserts

What it does:
 - Do not reset vy_stmt->n_upserts counter after scheduling squashing
   (when n_upserts is set to VY_UPSERT_INF).
 - In vy_squash_process(): apply statements that are inserted while we
   are reading an on-disk run. This is a dirty hack. Instead, we should
   fix the read iterator to handle this case.
 - Log errors if vy_squash_process() fails.

- Remove key_def from vy_read_task, use part_count instead.
- Fix handling of memory errors

Remove vy_stmt_info and construct vy_stmt directly from xrow. Need to
optimize read task.

No semantic changes.

- Move key stored as vy_stmt to vy_page_info.
- Remove vy_buf with min keys.

Issue #1945

Remove memory allocation from xlog_cursor_decompression. xlog_write_row
return xrow size.

Issue #1954

See #1610

- Use human-readable error messages instead of magic codes
- Use ER_COMPRESSION instead of OutOfMemory

- Handle xlog errors properly.
- Add error codes for decompression.

Issue #1972

* streamline upsert code base
* do reference couinting for allocated txn_stmt objects inside vinyl.c
* add column_mask optimization to upsert