feat(linalg): add AVX2+FMA SiLU f32 kernel

[cverrier]

x86 hosts with AVX2+FMA but no AVX-512 ran SiLU through the generic scalar kernel; they now get a vectorized one, which improves the throughput. Accuracy follows the FMA sigmoid f32 kernel.

Also, I benched the different SiLU f32 kernels on an AVX-512 machine:

cargo bench -p tract-linalg --bench silu -- silu_f32

Throughput results:

silu_f32/linalg         thrpt:  [1.8691 Gelem/s 1.8697 Gelem/s 1.8703 Gelem/s]

silu_f32/linalg-asm-fused
                        thrpt:  [3.0071 Gelem/s 3.0179 Gelem/s 3.0314 Gelem/s]

which gives a nice ~61% increase in throughput: this was expected because the x86_64_avx512_silu_f32_16n kernel is composed, not fused — it calls avx512_sigmoid_f32 directly. Therefore, on AVX-512 — which is backward compatible with AVX2+FMA — we can fall back to this new AVX2+FMA SiLU kernel for now; see #2567, where this change will be implemented.

Later, I will also implement a fused version for AVX-512.

[github-actions]

⚠️ Bench vs main — no speed regressions · 2 secondary regression(s)

Reference: 2026-08-02 morning nightly run (1d old) · full report → run

Speed — evaltime · prefill · decode

no inference-speed regressions

Improvements

Δ	metric	device	main → PR
🟢 -19.2%	parakeet_tdt_600m_v3_f32f32_decoder_pass evaltime · cpu	i9-11900kb_rtx-4060	0.838 ms → 0.677 ms

⚠️ 2 secondary regression(s)

Δ	metric	device	main → PR
⚠️ +8.2%	llama_3_2_3B_instruct_q40ef16_541 load+optimize · cuda	jetson-orin-nx	4.14 s → 4.49 s
⚠️ +5.9%	qwen3_1_7B_q40ef16_541 load+optimize · cuda	jetson-orin-nx	3.77 s → 4 s

[cverrier]

@kali A few checks got cancelled, looks unrelated to the PR itself. Could you re-run them when you get a chance?

[kali]

⚠️⚠️⚠️ Just rebased! ⚠️⚠️⚠️

[github-actions]

⚠️ Bench vs main — no speed regressions · 5 secondary regression(s)

Reference: 2026-08-04 morning nightly run (0d old) · full report → run

Speed — evaltime · prefill · decode

no inference-speed regressions

⚠️ 5 secondary regression(s)

Δ	metric	device	main → PR
⚠️ +20.9%	speaker_id RSS @ ready · pulse8	cortex-a7	17.7 MB → 21.4 MB
⚠️ +19.4%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a7	124 ms → 148 ms
⚠️ +19.3%	arm_ml_kws_cnn_m load · pass	cortex-a7	83 ms → 99 ms
⚠️ +19.3%	arm_ml_kws_cnn_m load · pass	cortex-a9	83 ms → 99 ms
⚠️ +14.2%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a9	127 ms → 145 ms

[cverrier]

@kali Is this work ready to be merged?