Check ev_async_pending() rather than state of the pipe to find
out whether we need a syscall. On a simple read-only test pipe->pipe
is empty while there is already a pending event more than a few times a
millisecond.

At the same time, checking ev_async_pending() is more expensive, since
it's an atomic variable, but overall performance increased
after the patch, so the cost seems to be justified.

Use staged mutex lock/unlock, cool off the WAL mutex

Since we no longer fork to make a snapshot, we don't need
an atfork() handler in box. We still need atfork() handler
in the app server to handle background=true option.

See also gh-517

refactoring: extract wal writer, rename replication.[hc] -> relay.[hc]
Extract the write ahead log into its own C file, rename
replication.[hc] to relay.[hc]
Cleanup headers.
No semantical code chagnes.
wal: begin changes necessary to switch to the common bus

Implement a generic inter-cord communication bus and use it
in separating network I/O and transaction processor.

Split all request processing work into network thread part
and transaction processor thread part, and make these
parts multi-threading safe.

Generalize fiber pool into an own data structure and API.

Misc cleanups.

Conflicts:
	src/iproto.cc

Conflicts:
	src/box/iproto.cc

Conflicts:
	src/iproto.cc

When selecting the next buffer to flush, take into account not
just its size, but how much of it has already been flushed.
This saves a couple syscalls per connection per event loop.

Refactor iobuf and extract ibuf/obuf into small allocators
library.
Fix leaking obuf_destroy() (some slabs could be leaked by
destroy).

Introduce rename to follow the style of small allocators library.

Display how much memory is used by a fiber, and the total size.

port_buf has similar purpose but uses mempool instead of realloc()