need be returned.  We keep a heap of the current best N tuples and sift-up
new tuples into it as we scan the input.  For M input tuples this means
only about M*log(N) comparisons instead of M*log(M), not to mention a lot
less workspace when N is small --- avoiding spill-to-disk for large M
is actually the most attractive thing about it.  Patch includes planner
and executor support for invoking this facility in ORDER BY ... LIMIT
queries.  Greg Stark, with some editorialization by moi.