Since we wrote `aes/matrix` to be easy to understand alongside the NIST paper rather than performant, I suspect there may be low-hanging fruit here.