In parallel replication, when the IO thread switches relay log,
the SQL thread re-opens the current relaylog and seeks to the
current position. There was a race that would cause it to
sometimes seek to the wrong position, causing corruption and
crash.