the float8 versions of the aggregates, which is all that the standard requires.
Sergey's original patch also provided versions using numeric arithmetic,
but given the size and slowness of the code, I doubt we ought to include
those in core.