OR MISS COMPLETE TRX

Problem:
=======

The SHOW SLAVE STATUS command contains the column RETRIEVED_GTID_SET.
This is supposed to contain the set of GTIDs that exist in the relay
log. However, the field is updated when the slave receiver thread
(I/O thread) receives a Gtid_log_event, which happens at the beginning
of the transaction.

If the I/O thread gets disconnected in the middle of a transaction,
RETRIEVED_GTID_SET can contain a GTID for a transaction that is only
partially received in the relay log. This transaction will
subsequently be rolled back, so it is wrong to pretend that the
transaction is there.

Typical fail-over algorithms use RETRIEVED_GTID_SET to determine which
slave has received the most transactions to promote the slave to a
master. This is true for e.g. the mysqlfailover utility.

When RETRIEVED_GTID_SET can contain partially transmitted transactions,
the fail-over utility can choose the wrong slave to promote. This can
lead to data corruption later.

This means that even if semi-sync is enabled, transactions that have
been acknowledged by one slave can be lost.

Fix:
===

It was implemented a transaction boundaries parser that will give
information about transaction boundaries of an event stream based on
the event types and their queries (when they are Query_log_event).

As events are queued by the I/O thread, it feeds the Master_info
transaction boundary parser. The slave I/O recovery also uses the
transaction parser to determine if a given GTID can be added to the
Retrieved_Gtid_Set or not.

When the event parser is in GTID state because a Gtid_log_event was
queued, the event's GTID isn't added to the retrieved list yet.
It is stored in an auxiliary GTID variable.

After flushing an event into the relay log, the IO thread verifies if
the transaction parser is not inside a transaction anymore (meaning
that the last event of the transaction has been flushed).

If transaction parser is outside a transaction, the I/O thread
verifies if a GTID was stored in the start of the transaction, adding
it to the retrieved list, ensuring that all the transaction has arrived
and was flushed to the relay log.

Also, before this patch, after the I/O thread flushed a single received
event into the relaylog, it was possible to rotate the relaylog if the
current relaylog file size exceeded max_binlog_size/max_relaylog_size.
After this patch, when GTIDs are enabled we only allow this rotation by
size if the transaction parser is not in the middle of a transaction.

Note: The current patch removed the changes for BUG#17280176, as it
      also dealt with similar problem in a different way.