If a mtr test case has started two mysqld processes (replication tests),
then kills the first one and kills the second one before starting the
first (so at some point there are two mysqlds down), then the ./mtr
waiting process bricks and forgets to monitor the "expect" file of the
first mysqld, so it never gets started again, even when its contents is
changed to "restart".

A victim of this deficiency is at least galera.galera_gcache_recover.

The fix is to keep a list of all mysqlds we should wait to start, not
just one (the last one killed).