A fix for https://bugs.launchpad.net/tarantool/+bug/1168076

No test case since the bug only manifsets itself under
concurrent load.

With addition of on_disconnect triggers, iproto_session_destroy()
was invoked from two places: the request handler code and
iproto_session_disconnect callback.

This could destroy the same session twice. Worse yet, it would
be put to the session cache twice, as a free session, and then
used twice for two different client sockets.

This could end up in an unpredictable mess, best case: libev
event loop runs infinitely since libev fd has events, while
session->input.fd is already set to -1 any attempt to use the
session ends up with an exception. At the same time the session
in this condition can not be destroyed since it's impossible to
tell libev forget about a file descriptor about which we already
know nothing.