Currently, if any Incident event occurs that breaks RPL, the incident
is binlogged without GTIDs. Due to this, if user wants to skip the
Incident event through general commands of GTID i.e.
'set gtid_next=<GTID>; begin; commit;', user cannot. User has to take
approach of setting the relaylog file and relaylog positions to post
incident_log_events to skip such incidents. Specially, when
autoposition is ON, user has to stop the autoposition, then set the
relaylog file/pos and then set it back to ON.

To fix the problem, we write the Incident event into stmt_cache, so
that a GTID is generated and written for it prior to flushing the
stmt_cache, and make the slave applier work with the change. Then
users can skip the Incident event through general commands of GTID
i.e. 'set gtid_next=<GTID>; begin; commit;'.

Remove testing purpose code at sql/parser.cc, since we don't
have a real case that generates an incident log event before
writing the real event to the binary log.

rpl_skip_incident.test and rpl_incident.test are removed, since
the related testing purpose code at sql/parser.cc is removed and
skipping incident on slave is tested by the two new added test
when gtid mode is on and by rpl.rpl_stm_binlog_max_cache_size
when gtid mode is off.

binlog_incident_ignore.test and binlog_incident.test are updated
with an incident caused by partially failed GRANT statement.