https://bugs.launchpad.net/tarantool/+bug/1200228
Remove the regression I introduced when trying to be to smart when
comparing integer numbers.

Numeric comparison is not memcmp compatible on x86, and I don't
know what I was smoking when I wrote a test which proved it wrong.

Save tuple field offsets for indexed fields in the tuple.
Merge tuple validation and initialization of field maps in the tuple.
Make formats work for an incorrect tuple.
Make tuple_field() inline.

Use formats when creating tuples.
Store format id in the tuple.

Refactor space cache code, create indexes
in space_create().

Remove init_storage option.

Move struct key_def into an own header.
Rename key_def methods to follow the coding style.
Rename field_data_type to field_type.
Instead of panicking on key_def initialization
error, throw an exception.

The allocator consists of a common slab cache, region and slab
allocators.
The patch provides unit tests.

It also modifies rlist_shift to be more in line with the
rest of rlist functions.

Before introducing a stronger row handler API,
eradicate tbuf from log_io/recovery, and pass around
a pointer to row and row length instead.

Extract the remaining places which access tuple->data
directly into tuple_to_port.cc file, which will have full
access to all tuple formats and all destinations to which
a tuple may need to be converted.

In this file it'll be possible to perform a conversion
efficiently, at the same time, it will be the only place
of cross-dependency between all tuple formats and all
conversion destinations.

- simplify the comparison function, only shortuct the single-part
  NUM key case.
- tweak mh_find() declaration
- remove wrappers around hash values
  and keys