Improve m68k __mulsi3

As part of some other work I was doing looking at constant multiplies on various platforms, and that led to looking a the em68k __mulsi3 code.

Which in turn led to the following microoptimizations. Note that this is desk-checked but UNTESTED, thus the RFC nature of the patches.

A bit of cleverness uses the available A0 and A1 registers to save three instruction words one memory fetch, shaving 16 cycles off the code.