Montgomery multiplication improvements

[fjarri]

• Added benchmarks for Montgomery multiplication and DynResidueParams creation
• Replaced the final reduction in montgomery_reduction() with sub_mod_with_hi(). Speeds it up by ~20% (for U256), I guess due to better vectorization?
• Extracted muladdcarry() from montgomery_reduction()
• Simplified Uint::sub_mod(), sub_mod_special(), add_mod(), add_mod_special() by reusing existing methods.
• Made DynResidueParams::new() a const fn
• Optimized mod_neg_inv calculation in DynResidueParams::new(), speeds it up by ~10% (for U256).

Originally I wanted to implement Montgomery multiplication by simultaneous multiplication + reduction instead of mul_wide, but it showed exactly the same performance, with the exception of the last reduction part - that's how I discovered the performance improvement. Still not quite sure what causes it.

(By the way, I measured the performance on arm64 - would be interesting to see the results for x64)

[tarcieri]

Seems like a nice simplification. Thanks!