Marginal performance improvement in LockBuffer --- calculate address

of BufferLocks[] entry just once.  Seems to save 10% or so of the
routine's runtime, which'd not be worth worrying about if it weren't
such a hotspot.