osd: pg: be more careful with locking around forced pg recovery

osd: pg: be more careful with locking around forced pg recovery

This does several little things that add up to big concurrency and safety
improvements:
* Switch to passing around PGRefs instead of raw pointers, which is
  generally a good idea
* drop the pg_map_lock once we're done looking up the PGRefs, since
  we don't need it and holding the PG pointer alive was the only previous
  thing that might have made it necessary
* don't hold the recovery_lock since we don't need any OSD-level
  synchronization
* make sure the PG is not being deleted before we do a force-change of its
  state

Fixes: http://tracker.ceph.com/issues/20808
Signed-off-by: Greg Farnum <gfarnum@redhat.com>