k_mul() when inputs have vastly different sizes, and a little more
efficient when they're close to a factor of 2 out of whack.

I consider this done now, although I'll set up some more correctness
tests to run overnight.