[fft] AVX2 Shoup butterflies with runtime dispatch

[devin-ai-integration]

Summary

Vectorizes the hot NTT loops for 32-bit prime moduli using Shoup multiplication (precomputed quotients w' = floor(w·2^32 / MOD) stored alongside rt/inv_rt), avoiding Montgomery form entirely: data stays in the normal residue representation, lazily in [0, 2·MOD) between passes and normalized back to [0, MOD) on a transform's final pass — outputs are bit-identical to the scalar path.

New AVX2 kernels in fft_core (gated by ntt_simd<num>, true for modnum<MOD> with MOD < 2^30 on x86):

• forward_simd / inverse_simd: butterflies at k >= 8 are 8-wide (mul_shoup = mulhi32 + 2 mullo, lazy_reduce = min_epu32(x, x - lim)); the k < 8 tail stays scalar Shoup.
• scale_simd: constant-multiply, used by fft_engine::finish's 1/n scale for assign ops (finish_linear regains an assign fast-path so the cut fixup doesn't force the scalar cut_op route).
• twiddle_simd: the b[i] = coeffs[i] * rt[n+i] pass in extend.

The kernels carry [[gnu::target("avx2")]] and are selected at runtime via __builtin_cpu_supports("avx2"), so no special compile flags are needed and non-x86 / non-AVX2 / other value types (goldilocks, cnum, dynamic modnum) keep the existing scalar loops untouched.

Measured (g++ -O2, no arch flags, N=2^18): isolated butterflies ~0.58 → ~0.20 ns/weighted-op (~2.8×); end-to-end multipoint eval 325 → ~233 ms (−29%), interpolation 455 → ~337 ms (−26%). Remaining scalar hot loops are the variable×variable pointwise mul/mul2 passes, which Shoup can't cover (they'd need Montgomery-domain transforms).

Link to Devin session: https://app.devin.ai/sessions/44c622b53d88445c9431863199c4fe7b
Requested by: @ecnerwala

---

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[github-actions]

GCC Code Coverage Report

📂 Overall coverage

Metric	Coverage
Lines	🟡 7595/9420 (80.6%)
Functions	🟡 987/1106 (89.2%)
Branches	🟡 6177/7925 (77.9%)

[beta-devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.