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.