Minor performance improvement in transition to external sort

The attached patch replaces the existing siftup method for heapify with
a siftdown method. Tested with random integers it does 18% fewer
compares and takes 10% less time for the heapify, over the work_mem
range 1024 to 1048576.

Both algorithms appear to be O(n) (contradicting Wikipedia's claim
that a siftup heapify is O(n log n)).

--
Cheers,
    Jeremy