~/.config/tarantool superseeds /etc/default, not augments it

Conflicts:
	src/CMakeLists.txt
	src/lua/init.cc
	src/lua/log.lua

Don't run event loop if socket is ready to write. *Always* write
optimistically. With fiber.channel(), channel push  wakes up the waiter
without in the same event loop iteration. This means that whenever we have
data on an incoming connection, we can push messages to outgoing connections
in the same iteration, without epoll_ctl syscalls or unnecessary
epoll_wait.
With 1 incoming and 5 outgoing connections, for example, this saves
~10 syscalls per each incoming request.

If a fiber becomes ready on a user-defined event, schedule it
in the same event loop, not in the next loop, to save one epoll_wait()
invocation in cases when something is added to a fiber channel,
and there is a fiber ready for execution waiting on the channel.

Add a micro-bench to the unit test suite.

Rollback d29469f2, it makes build bug on FreeBSD build and doesn't affect on OSX

Conflicts:
	doc/sphinx/book/configuration.rst
	doc/sphinx/book/connectors/__python.rst
	doc/sphinx/book/connectors/index.rst

Change fiber_sleep(0) implementation to use fiber_wakeup().
Implement a more fair scheduling for "ready" fibers: they are
scheduled in the next event loop iteration, not in the same.

Fix a bug with readchunk returning more than was requested.

Fix mistype in load_cfg.lua (local variables)