- The server side of the DLM wasn't always handling the invalid handle case

very gracefully.  Since there are races that make invalid handles not hugely
uncommon, that's fixed now.
- Even on the client side, we could get cancelled between ldlm_lock_decref() and
ldlm_lock_cancel(), thus passing an invalid handle into cancel().  This is
handled now.
- Fixed another leaked request.
- Fixed an IT_SETATTR bug, even though we don't currently exercise that code
path.
- Fixed common.sh for the multiple mount cleanup case
- llcleanup.sh was unloading things in the wrong order (MDS is pinned by LDLM
and needs to be unloaded later in the process).  Fixed.