BoxedUint: use Limbs in montgomery_mul

[tarcieri]

This was previously written using raw Words to ensure maximum performance, however it precludes access to the constant-time helper functions implemented on Limb.

This carefully rewrites the implementation to using Limb instead of Word while avoiding any performance impacts.

Benchmarks (unchanged)

Montgomery arithmetic/modpow, BoxedUint^BoxedUint
                        time:   [24.424 ms 24.444 ms 24.465 ms]
                        change: [-0.1629% -0.0399% +0.0760%] (p = 0.52 > 0.05)
                        No change in performance detected.

[tarcieri]

@dignifiedquire I suspect what was hitting us before were the places there are currently branches, namely the comparisons between words/limbs. I suspect that subtle has a rather slow constant-time impl of integer comparisons which made things massively slower.

The data-dependent branching still needs to be addressed but I have some ideas.

[dignifiedquire]

cool, I am glad this is possilbe