WASM PR 2: move the WASM microbenches out of src and into the bench target

[czoli1976]

Stacked on #2563 — its two commits are the first two here.

The wasm backend kept three timing modules inside src/ behind #[cfg(test)] + #[ignore], which no other backend does. benches/wasm.rs already supersedes two of them, and says so itself: its header records that the in-src version looped all four GEMV kernels back-to-back and so biased whichever ran last. Those two go; the activation bench moves across to join it.

numerical_consistency_16x1_vs_32x1 stays in src, in wasm/dispatch_tests.rs — it is an assertion rather than a timer, and the only numerical cross-check between two kernels.

The four copies of the kernel-lookup helper in benches/wasm.rs collapse to one. I left the timing loops alone: they use 50 versus 200 warmup calls and only two of them do min-of-N reporting, and those differences are deliberate, so a shared helper would have quietly changed what the numbers mean.

Net: 335 lines out of src, and wasm.rs is down to 78 lines.

Nothing here ships — test and bench code only, and the kernel disassembly is unchanged. The bench runs: on an M1 Pro the relocated activation bench reports sigmoid 1.49–1.54x and tanh 1.40–1.62x over the scalar path, matching what the module documented.

🍍

[kali]

⚠️⚠️⚠️ Just rebased! ⚠️⚠️⚠️

[github-actions]

🔴 Bench vs main — 3 speed regression(s) · ⚠️ 7 secondary

Reference: 2026-08-03 morning nightly run (0d old) · full report → run

Speed — evaltime · prefill · decode

Δ	metric	device	main → PR
🔴 +15.8%	hey_snips_v4_model17 evaltime · 2sec	cortex-a9	38.7 ms → 44.8 ms
🔴 +10.7%	inceptionv3 evaltime · pass_mt	cortex-a53	1.12 s → 1.24 s
🔴 +8.6%	hey_snips_v4_model17 evaltime · 2sec	cortex-a7	43.6 ms → 47.3 ms

Improvements

Δ	metric	device	main → PR
🟢 -22.0%	parakeet_tdt_600m_v3_f32f32_decoder_pass evaltime · cpu	i9-11900kb_rtx-4060	0.838 ms → 0.654 ms
🟢 -11.1%	hey_snips_v4_model17 evaltime · 2sec	apple-m1-max	0.979 ms → 0.87 ms
🟢 -8.0%	hey_snips_v4_model17 evaltime · 2sec	i9-11900kb_rtx-4060	0.972 ms → 0.894 ms
🟢 -6.8%	inceptionv3 evaltime · pass	apple-m1-max	63.5 ms → 59.2 ms
🟢 -6.7%	hey_snips_v1 evaltime · 400ms	apple-m1-max	0.13 ms → 0.121 ms

+2 more improvement(s)

Δ	metric	device	main → PR
🟢 -6.3%	hey_snips_v4_model17 evaltime · pulse8	apple-m1-max	0.117 ms → 0.11 ms
🟢 -4.6%	openelm_270M_q40ef16_541 decode · cpu	apple-m1-max	12.7 ms/tok 79 tok/s → 12.1 ms/tok 82.82 tok/s

⚠️ 7 secondary regression(s)

Δ	metric	device	main → PR
⚠️ +15.5%	arm_ml_kws_cnn_m load · pass	cortex-a9	84 ms → 97 ms
⚠️ +15.2%	hey_snips_v4_model17 load+optimize · 2sec	cortex-a9	4.39 s → 5.06 s
⚠️ +14.8%	hey_snips_v4_model17 load+optimize · 2sec	cortex-a7	5.07 s → 5.82 s
⚠️ +14.1%	hey_snips_v4_model17 load · 2sec	cortex-a9	3.81 s → 4.34 s
⚠️ +13.4%	hey_snips_v4_model17 load · 2sec	cortex-a7	4.34 s → 4.92 s
⚠️ +9.2%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a9	131 ms → 143 ms
⚠️ +5.7%	llama_3_2_3B_instruct_q40ef16_541 load+optimize · cuda	jetson-orin-nx	4.14 s → 4.38 s