Not being transactional, View Change events always suffered with
problems when we wrote them to the binlog and relay log.

On Relay log:
When the View Change event was the last event in the
log, this meant that the relay log positions wouldn't be update.
Upon plugin restart this meant that the events would be again written
to the binlog.

On Binlog:
When written to the binlog the events would not any support to allow
skipping them when we connected to the same master twice for example.
Other problems were that we are not able to track when the event was
written to relay_log/binlog without checking log positions or thread
states.
MTS issues are also predictable in the near future

Solution:
When queued to the relay log, the View Change event will be surrounded
by a transactional block.

GTID
BEGIN
VIEW CHANGE
COMMIT

This need changes to the event writing routines as it now needs to be
written to the binlog cache that will be flushed by the commit.