Extend atomic DDL to make replication crash-safe for DDL

Atomic DDL makes stand-alone DDL crash-safe. But there are still a couple issues pending for making replication completely crash-safe.

1. Sometimes, the DDL must be binlogged with a specific GTID. Either on the slave, replicating an existing GTID, or on the master, replaying mysqlbinlog output (or otherwise user-specifed GTID). The DDL recovery/log replay must save any such GTID in the DDL log and apply it if binlogging during recovery.

2. The mysql.gtid_slave_pos table must be updated with the replicated GTID if a crash happens on the slave during DDL, and such DDL is completed by DDL recovery.

Monty committed a patch for the DDL-side of this, see 6ea6864185419a6d52da6115e4d0862fed5af7ef