each tuple, as per my proposal of several days ago.  Also, clean up
sort memory management by keeping all working data in a separate memory
context, and refine the handling of low-memory conditions.