yet another approach to fix the loop lock order inversions v5

[Christoph Hellwig <hch@lst.de>]

Hi all,

this series uses the approach from Tetsuo to delay the destroy_workueue
call, extended by a delayed teardown of the workers to fix a potential
race window then the workqueue can be still round after finishing the
commands. 

Changes since v4:
 - keep the (questionable) __invalidate_device call in nbd as-is for now
 - suppress uevents while reconfiguring

Changes since v3:
 - change bd_openers into a atomic_t, including a bunch of cleanups
   and fix found while adding those

Changes since v2:
 - rebased to the lastest block for-next tree, which has the async
   clear reverted and ->free_disk
 - impkement ->free_disk for loop to handle open vs delete races
   more gracefully
 - get rid of lo_refcnt entirely

Changes since v1:
 - add comments to document the lo_refcnt synchronization
 - fix comment typos