Move the  block with @finally into an own function.

Implements
https://blueprints.launchpad.net/tarantool/+spec/fiber-specialization

After this patch, all requests even within the same session
are handled asynchronously, and responses are sent to the client as
soon as they become ready, regardless of the request position
in the input buffer.

This also changes  the internal implementation of iproto
from a "one fiber per connection" pattern to "fiber work crew"
pattern. This approach significantly reduces the amount of
context switches between fibers.

There is a bug in gcc Objective C support and nested exception,
with which an exception passes by a @catch clause if
there are nested @finally clauses. This patch doesn't work
around this bug, so if the server is compiled with gcc and
the client closes its side of the socket without fully reading
the input, the server terminates. The server needs to be compiled
with clang for now to work around this bug.

Do not modify output buffer between yield points:
in net-split, multiple fibers work on the same
session, and append output to the same output
buffer.

Fix argument getpeername

Change sio_advance_iov to never modify iov vector in place.
This allows for reuse of this function in a raw event
loop, without fibers.

Minor cleanups.

Make memcached/noreply.test less race prone:
flush all command is running in the event loop, and sometimes
it doesn't get a chance to finish before we issue the next
command.

Use a large timeout value to indicate an infinite timeout,
not 0. Chris J. Date teaches us that using 0 as a special
value of infinity is a bad idea.

An incomplete fix for a bug when a channel is destroyed
in atexit handler, but has fibers blocked on it.

Minor cleanups.

Get rid of an artificial wrapper around libev's ev_io object.
This allows for more code reuse.

 * join fiber->ready and fiber->ifc into one fiber->state
 * create fiber_yield_timeout function
 * channel uses fiber_yield_timeout
 * ipc test doesn't use fiber.sleep

Move memcached to src/
Make iproto.m dependend on box/*, not vice versa.
Remove box_ prefixes from test suite names.
Initialize networking outside of box/.

This makes Box a step closer to being just a storage
layer for Tarantool.

Still some ugly deps since box/ continues to manage
local hot standby and memcached expire (todo: move it out).

coio_writev() could modify the passed parameter in case of a
partial write.

It would corrupt struct obuf, which would pass to coio_writev()
a pointer to a vector whose iov_base pointers were later reused.

The following would happen:
- iobuf_flush() would pass obuf->iov to coio_writev
- coio_writev would advance one or two iov_base pointers
- capacity value would hence become meaningless, and a subsequent
obuf_dup() would overwrite heap memory
- we would get a memory corruption in some place very distant
from where the problem originally occurred.

The fix is to preserve iovec values in coio_writev.
This is implemented by keeping a copy of the changed iovec
and restoring the original iovec value at exit from
coio_writev.