Explicitly close net.box connection on error.

Partially reverts squashed 316d4e3a

This regression introducing during replacing port_ffi with port_buf
by 2fcce449

iproto thread can add quirks to thread scheduling, replace
single-call yields with yields in a loop.

Correct return value of ipc_channel_put(), since
it may return an error if the channel is closed.

Make get/put functions inline.

Remove a public declaration for ipc_channel_is_readonly() (towards
removing the concept of readonly altogether).

after review fix: remove group name param in config.

gh-918 tarantoolctl group support implemented with default user group

This test was disabled by me during 1.6 async-master-master development.
Since Tarantool doesn't use proc_title for replication status anymore,
I added a new test that check box.info.replication.

Prepare for multi-master.

Linux-specific option obsolete since kernel 2.2

Wait() no longer called hence waiters not initialized. Does it make
sence to have a test for :waiters (undocumented internals)?

- Waiting on a brand new socket (neither listen nor connect called)
  completes immediately with HUP status on Linux. However on FreeBSD/osx
  it never completes.

- Attempting to connect to a port bound on a loopback interface which is
  not yet listening yield ECONNREFUSED in blocking mode on Linux. On
  FreeBSD/osx blocking connect never completes ("never" means it doesn't
  complete in a reasonable time).

- Calling shutdown() on a listening socket works in Linux but in
  FreeBSD/osx it doesn't.

Note: it was incomplete, made thing unnecessary complex and it did not
align with the CMake way (c) for crosscompilation.

Pass sysroot on OSX.
Make the code in luajit.cmake a bit more sane and straightforward.

Note: v1 of this patch was reverted due to performance regression.

* log messages correctly in case of SocketError
* a step towards making replica_connect() reusable
  from the outside.