Use partial-product multiplier for low-latency FMA

The area-optimized (i.e. speed = false) version doesn't use the PPM;
it uses behavioral A*B+C then does LZA afterwards.