Daniel Blueman reported a bug with fio+balance on a ramdisk setup.
Basically what happens is the balance relocates a tree block which will drop
the implicit refs for all of its children and adds a full backref.  Once the
block is relocated we have to add the implicit refs back, so when we cow the
block again we add the implicit refs for its children back.  The problem
comes when the original drop ref doesn't get run before we add the implicit
refs back.  The delayed ref stuff will specifically prefer ADD operations
over DROP to keep us from freeing up an extent that will have references to
it, so we try to add the implicit ref before it is actually removed and we
panic.  This worked fine before because the add would have just canceled the
drop out and we would have been fine.  But the backref walking work needs to
be able to freeze the delayed ref stuff in time so we have this ever
increasing sequence number that gets attached to all new delayed ref updates
which makes us not merge refs and we run into this issue.

So to fix this we need to merge delayed refs.  So everytime we run a
clustered ref we need to try and merge all of its delayed refs.  The backref
walking stuff locks the delayed ref head before processing, so if we have it
locked we are safe to merge any refs inside of the sequence number.  If
there is no sequence number we can merge all refs.  Doing this not only
fixes our bug but keeps the delayed ref code from adding and removing
useless refs and batching together multiple refs into one search instead of
one search per delayed ref, which will really help our commit times.  I ran
this with Daniels test and 276 and I haven't seen any problems.  Thanks,

Reported-by: Daniel J Blueman <daniel@quora.org>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>