The wait_for_commit mutex is reused by parallel replication for transactions
that are completely unrelated (as part of the rpl_group_info freelist).

This can cause SAFE_MUTEX to give false warnings about incorrect locking
orders, where first M1 is taken while M2 is held, and later M2 is taken while
M1 is held.

The fix should be to reset the mutex when it is reused for a new purpose, in
SAFE_MUTEX builds.