Draft: pan/mdg: Experiment with masks at 16-bit granularity

I was curious how much we were being hurt by dealing with 8-bit masks when we can get away with 16-bit masks instead without loss of generality. Turns out -- not too much, surprisingly. Was a useful experiment and will help inform some stuff for Bifrost design, but overall only about a 1% reduction in compile-time, boo.

Marked draft because the RA change is causing some weird looking spills on manhattan that I haven't figured out yet. Said spills might be hiding a bigger win in compile-time as well.