perf: faster Fp.add/sub with branchless reduction

[MatteoMer]

Problem

Fp.add at 4.6-5.7ns vs arkworks' 3-4ns (~40% gap). The conditional modular reduction likely causes branch misprediction overhead.

Approach

Replace conditional branch with branchless reduction:

• x86 ASM using sbb to generate mask + cmov chain for conditional subtraction
• Or: always subtract modulus, use carry flag to select result (sbb + cmov)
• Ensure no unnecessary loads/stores around the operation

Impact

Add/sub are called even more frequently than mul (every field operation involves adds for Karatsuba cross-terms, point coordinate updates, etc.). A 30% improvement compounds significantly.

Files

• src/field/mod.zig — MontgomeryField.add(), .sub(), .addNoReduce(), .reduce()

Benchmark baseline

• Fp.add: 4.6-5.7 ns/op
• Target: 3-4 ns/op