Implement GTID "strict mode"

Current GTID implementation allows situations when slave is not configured read-only and accepts any transactions from clients. These transactions get inserted in the middle of replication stream and then don't go anywhere. So essentially they generate data drift without any error messages and without easy methods of detection.

To make production environments more bullet-proof there should be a flag turning on the "strict mode" for GTID replication in which slave accepts from master and binlog allows to commit only transactions with strictly increasing seq_no for each domain.