Move out from box.m INSERT/UPDATE/SELECT/DELETE/CALL
implementation.

Cherry-picking code from txn-refine.
Remove an extra member in struct box_txn.
Rename txn->tuple to txn->new_tuple.

Fix bug https://bugs.launchpad.net/tarantool/+bug/990728
"box.space:truncate() is a CPU hog".

The old truncate algorithm would restart iteration
from the beginning to delete every tuple.
With HASH index it was especially slow, since
it would have to skip more and more empty slots
to find a non-empty slot for deletion.

Rewrite the truncate() to walk linearly
over all elements and delete them.

This works both with HASH and TREE indexes.

On the same token make truncate multipart-index
aware (was missed in multipart-pk patch).
Add a test for that.

To work with multipart keys efficiently,
implement tuple:slice(). A slice returns
a linear set of fields from a tuple.
Convenient to retreive a multipart key or
tuple tail. Document tuple:slice().

Convert box.lua to box.lua.c and compile it in.
Fix broken out-of-source build of luajit, by adding
a few more generated files to touch-nocreate rule.

Replacing salloc() with malloc() found a bug triggered
by multipart key patch:
we don't always check for key cardinality.

When the slab alloc arena is fully used, and we try to
truncate a space, we could crash, since

a) we never check return value of salloc()
b) we used to salloc() an iterator for truncate(), and
it would return NULL, and we would try to access it without a check.

The fix is:

- to throw exceptions directly from salloc()
- to not use salloc() for iterators, since we still should
be able to truncate a namespace when the slab alloc arena is
full.

'find' TreeIndex class member was fixed. Now allocate enough space for key_data.

Spelling fixes in the error message file.

Add more information to the error when key cardinality
doesn't match index cardinality in exact match.

Avoid server start/stop in box_big/hash.test (ugly,
but until we have named/dynamic spaces, which are already in the
pipeline, we should not make our test procedure slower).

Rename multipart-pk.test to tree_pk_multipar.test:
consistently use underscores, not dashes, in file
names.

(Unrelated) Recycle more error codes.

Make keyValue an inline function to save a few CPU cycles.

Avoid creation of an extra table in box.lua just to find out
variable list array cardinality.

Revert changes to box.update():
a) they break backward compatibility

This reason alone should be sufficient. There must be no changes
which break backward compatibility in a stable release, unless
it's a fix for a crashing/security bug.

box.update() changes got to be done in a separate task,
and should not break backward compatibility.

Leave update_ol for now, to test multi-part primary key updates,
even though the function itself is quite heavy, hardly something
we can recommend while running Lua procedures in production.

Warnings was removed. Minor coding-style cleanup.

'init.lua' start-up file loading was moved from box.lua to tarantool
initialization section after main modules initialization (lua, mod, admin,
etc..). New tests for 'init.lua' was added.

This reverts commit 8dc70aac.

Minor code clean-up.

Make sure we do not try to destroy uninitialized
mutexes in the snapshot thread.

Implement FIFO policy for the wal writer queue.

Code review fixes:
     - Key validation procedure was moved to Index classes;
     - Additional hash tests was added.

This patch introduces wal_mode configuration variable,
which can take either "fsync" or "fsync_delay" values.

In "fsync" mode, we open the write ahead log in O_SYNC
mode, and thus sync every write to disk.

Th fsync_delay mode is the backward-compatible mode:
in this mode, we only do fsync once every fsync_delay
seconds. If fsync_delay is 0, there is no fsync.

Otherwise, this patch is only adding auxiliary headers,
object files, etc.