WL#1127: Optimize fatal_error and kill usage

The idea with this task is to make it faster and easier to detect is
something went wrong.

Currently we have to the following tests in sql_parse.cc:

in handle_one_connection:

    while (!net->error && net->vio != 0 && !thd->killed)
    {
      if (do_command(thd))
	break;
    }

This should be replaced with:

    while (!thd->killed)
      do_command(thd);

(In other words, we should set thd->killed both if do_command() would
otherwise return true, which should only happen on 'quit', or when
net->error or net->vio are set.  The later should never be set after
the loop has been entered.

In dispatch_command() we have:

 while (!thd->killed && !thd->is_fatal_error && thd->lex.found_colon)

This should be replaced with:

 while (!thd->killed && thd->lex.found_colon)

In other words, we should also set thd->killed when we set
thd->is_fatal_error.  This will make this and other errors easier to
handler.

The above should be relatively easy to do after dimitri's patch where
thd->killed can be used to only kill a running query.
(We probably have to add a new killed state for the case of 'fatal_error',
which should work like KILL_QUERY).