Problem
-------

The slave server is losing events not yet applied when
MASTER_AUTO_POSITION=0, both replication threads are stopped and the
applier delay is changed using CHANGE MASTER TO MASTER_DELAY.

Analysis
--------

WL#6120 introduced the possibility of issuing CHANGE MASTER TO commands
without having both replications threads started.

Issuing CHANGE MASTER TO RELAY_LOG_FILE/RELAY_LOG_POS/MASTER_DELAY is
possible with only the SQL thread stopped.

When both thread are stopped and a CHANGE MASTER TO MASTER_DELAY is
being issued, the slave is purging the relay log. In this case, as it is
possible that the SQL thread was not in sync with the I/O thread, the
SQL thread positions should be copied to the I/O thread. The I/O thread
will then start replicating from where the SQL thread stopped, but this
(the copy of the positions) is not happening.

Fix
---

Fixed the behavior of the MASTER_DELAY option.