The old patch was failing because sometimes the inodes were gone by the time

The old patch was failing because sometimes the inodes were gone by the time
the callback was called with CB_DYING.  I also realized that the client would
believe that it had valid data for too long.

Instead of an extra callback when the lock is freed, I moved that extra
callback to just before we cancel the lock.  Because there's two paths for that
now (cancelling with server notification and cancelling without), I had to
introduce _another_ l_flag, which indicates whether we've done this call.  Oh
well.  It works much better now.