modexp: 2.5x accel on small exponent

[mratsim]

Improve a potential DOS vector for the modexp precompile.
The overhead of computing r² was too high, changing to the reduce operation compared to repeated doubling (which is used for the compile-time proc because reduce requires pointers) improves perf by 2.5x.

Todo:

• Provide a vartime reduction to improve perf further
• Use it at compile-time to improve compilation speed
• Evaluate Barrett Reduction for small modexp instead of Montgomery
• Or potentially just multiplication and modular reduction

[mratsim]

Followed up by #286