M_STATUS == DA_ERROR || M_STATUS == DA_OK

Problem:
=======

An assert will fail when the replication sql thread on a slave server
is killed while waiting for commit lock will.

Analysis:
========

In the case of a failed commit attempt at
Xid_log_event::do_apply_event(), an error message will be reported
informing the reason why the commit could not be completed.

The error message to be reported is collected using message() at
sql_error.h that contains an assert expecting that current m_status 
must be DA_ERROR or DA_OK before returning the message.

The problem happens when the SQL thread is waiting for the commit lock
at MDL_context::acquire_lock and is killed.

In this case, the resulting wait_status will be MDL_wait::KILLED, and
will result in acquire_lock() returning an error but without throwing
one (both other cases, MDL_wait::VICTIM and MDL_wait::TIMEOUT, throw
errors using my_error()).

So, in the case of a thread being deliberate killed while waiting for
a lock would lead to a case where no error was flagged, but also no
success was flagged, making the assert expecting DA_ERROR or DA_OK to
fail.

Fix:
===

The code in MDL_context::acquire_lock in mysql-5.7 already throws an
error for MDL_wait::KILLED. This patch just back-ported this behavior.

Added also some debug sync points to easy putting sql thread in
"Waiting for commit lock" state.