Make sure that s_umount is acquired *before* we drop the final
active reference; we still have the fast path (atomic_dec_unless)
and we have gotten rid of the window between the moment when
s_active hits zero and s_umount is acquired.  Which simplifies
the living hell out of grab_super() and inotify pin_to_kill()
stuff.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>