Post-push refinement to fix a valgrind issue.
The basic pushed patch did not expect a block of
 if (!ret_worker->checkpoint_notified) {}
be executed multiple times during scheduling of one group.
In fact it should have been left intact to continue to be executed at T-event
time that guarantees its single time run.
And that is restored, that is the former patch part is reverted.

As the matter of fact the physical coordinates were unknown to the Worker 
only during time of the first transaction scheduling upon master binlog
rotation or the very first one at MTS start.
Information about the master binlog at such points is passed to the Worker
now via augmented notification mechanism.
The new notification is light as in implementation so in terms of execution
(is supposed to be pretty rare) and never more that once during one transaction
scheduling (as asserted in the DBUG version of the patch).

Attention to rpl_conflicts.test is paid to sort out 
rpl_{row,stm}conflicts.test failing on PB2.