After generating GTID for incident log event, skipping slave error
code 'ER_SLAVE_INCIDENT' does not work. The root cause is that we
did not commit the transaction to consume its owned GTID when
applying the incident log event in the case, and more the slave
job added by the incident's GTID is wrongly removed from the
global assigned queue when MTS is enable.

To make skipping slave error code 'ER_SLAVE_INCIDENT' work well,
commit the transaction to consume its owned GTID when applying
incident log event in the case and remove the slave job added
by the incident's GTID correctly when MTS is enable.
To make NDB cluster happy, if there is no binlog cache then we
write incidents into the binlog directly.

(cherry picked from commit 4585ae70078721d7ee1ed23ee632a251ac32de41)