After a fetch failure and stage retry, we may have multiple tasksets
which are active for a given stage.  A late completion from an earlier
attempt of the stage should update the most recent attempt for the
stage, so it does not try to submit another task for the same partition,
and so that it knows when it is completed.