MDEV-7458: Deadlock in parallel replication can allow following transaction to start replicating too early

In parallel replication, don't rollback inside ha_commit_trans() in case of
error.

The rollback will be done later, but the parallel replication code needs to
run unmark_start_commit() before the rollback to properly control the
sequencing of transactions.

I did not manage to come up with a reliable automatic test case for this, but
I tested it manually.