Add opt-in BMI2/ADX Montgomery multiplication for the Pasta fields on x86_64

[lamb356]

Summary

Adds an opt-in BMI2/ADX inline-assembly Montgomery multiplication backend for the Pasta fields on x86_64, behind a new x86_64-asm feature (off by default). The repo already ships an aarch64 assembly path for Apple silicon; this is the x86_64 counterpart.

• New crates/pasta_curves/src/fields/x86_64_asm.rs: 4x4-limb Montgomery multiply/square using mulx/adcx/adox carry chains, wrapped in narrow unsafe functions.
• Fp/Fq dispatch to the asm path for mul, square, and repeated square-and-multiply when feature = "x86_64-asm" and target_arch = "x86_64"; the portable Rust path is unchanged and remains the default everywhere else.
• No changes to representation, transcripts, proof format, or verifier behavior — this is a drop-in replacement for the field multiplication inner loop (~22ns → ~16.5ns per mul on our test box).

Measured on the Orchard k=11 prover benchmark (8 threads): roughly 10ms off the end-to-end proving time on its own. All numbers in this PR series were measured on an 8-thread VM — if you have a 16-core machine handy we'd appreciate your numbers, since the dominant stages scale with cores. Reproduce with:
RAYON_NUM_THREADS=16 ORCHARD_K11_PROVER_THREADS=16 cargo bench -p zakura-orchard --features circuit,orbits --bench orchard_k11_prover

Part of the split of #273 (piece 1 of 6, as proposed there).

Testing

• cargo test --release --features x86_64-asm in crates/pasta_curves (91 + 1 tests pass, exercising both fields against the portable implementation).
• Default build (no feature) unchanged; cargo check --locked --all-features --all-targets and MSRV (1.91) check clean.
• cargo fmt --all -- --check clean.

[ValarDragon]

Will test this same time were testing all the other x86 work in repo!