Description of problem:
    When add-brick is performed on a replica 2 subvolume that has a file in
    data split-brain, the current logic for post-op failure accounting
    (originally added to fix spurious cyclic split-brains) picks up one of
    the 2 bricks as source, heals it into the newly added brick and resets the
    afr xattrs on  all 3 bricks. But the brick that was not chosen as source will
    still have different contents leading to silent data mismatch. The same
    problem is there even for metadata split-brain.