Not-released memory after stop-slave was caused by skipping MTS cleanup
in case the Workers initialization fails.
It could fail for instance because of imcompatibility of
max of opt_mts_checkpoint_group and sizeof(mysql.slave_worker_info.Checkpoint_group_bitmap)
when the former allowed a value over the size (fixed by Bug@12979420).

Fixed with refining MTS stop procedure, in particular making sure that cleanup of
Coordinator related objects is done even though Workers might not have been threaded.

Also @@slave_checkpoint_group system var is made cacheble 
into rli->checkpoint_group at slave start same way as slave_parallel_workers.
That is a new slave session will use the last available value of @@slave_checkpoint_group
on the eve of START-SLAVE, and changes to the variable won't have any effect to the
running slave session.