tuple.cc passed wrong parameters to small_alloc_create - prealloc
size instead of slab size, and slab size instead of prealloc size.

When correct parameters were supplied, it turned out that
objsize_max is calculated wrongly, and the last factor pool
in small allocator is too framgented (uses 3 empty slots instead
of 4, with 25% overhead).

The problem would not show up in any test, only increase
memory fragmentation.

This patch also fixes this problem.

Conflicts:
	test/box/fiber.result
	test/box/fiber.test

Conflicts:
	doc/user/connectors.xml
	doc/user/language-reference.xml
	doc/user/stored-procedures.xml

Conflicts:
	CMakeLists.txt
	client/tarantool/query.c
	client/tarantool/tc.c
	client/tarantool/tc_cli.c

Move towards a uniform include scheme where header files are
stored along with the sources files, in the same directory.

We had to begin to use this scheme thanks to use of subprojects,
and a mix of two approaches turns the code base into a mess.

Move recovery to box/. Now only replication.cc and replica.cc
depend on recovery.

Remove unused headers from a bunch of files. This should reduce
recompile time and ease future refactoring.

slab_arena allows to have multiple allocation
sources of differnt type for different purposes.
E.g. one can have MAP_PRIVATE arena for runtime
memory and MAP_SHARED for tuples.

The arena can be shared across multiple threads.
Implement such arena, cover with tests, use
it for region allocators.

This is done to avoid situation, when two or more
fibers yield inside their try/catch blocks and
throws an exceptions. Seems like exception unwinder
stores some global state while being inside
a catch block.

This could lead to incorrect exception processing
and server crash.