Field optimization

[mratsim]

This does the following

1. Significantly improve point deserialization speed of Twisted edwards curves by implementing a generic over any prime, optimal fused sqrt+division. This is significantly faster than litterature p ≡ 5 (mod 8) only optimization uv³(uv⁷)^((p−5)/8) presented in Bernstein et al paper for ED25519.
   
2. Implement addition chains for inversion and sqrt of Curve25519. Note, Constantine is needlessly using Montgomery representation for the prime field. Montgomery representation is used to reduce the cost of modular reduction, but pseudo-mersenne primes in the form of 2ᵐ-c have a significantly more efficient more reduction. Modular multiplication cost could be made ~40% cheaper (Finite field computation for moduli of special form #11)
3. Rework square roots public/private proc, isSquare now uses addition chains (important for BLS12-377 sqrt on Fp2)