64b mults fail silently

[dpetrisko]

Currently we deschedule mulh so that GCC avoids them:
bespoke-silicon-group/riscv-gcc#3

But if users explicitly request it, it’s a confusing error. Or the compiler could generate in performance critical loops

1. We should put in an assertion so it fails in sim. This is non-controversial
2. hardware we could maybe trap
3. it’s also fairly cheap to put in an iterative mulh (Support MULH black-parrot/black-parrot#1110). Parameterization for this case seems icky because we’re technically non-compliant with M-extension, so I would vote for adding the hardware