look at the actual aggregate transition datatypes and the actual overhead
needed by nodeAgg.c, instead of using pessimistic round numbers.
Per a discussion with Michael Tiemann.