core: support tf_half_pixel_for_nn resize coordinate mode

[kali]

Resize rejected the legacy tf_half_pixel_for_nn coordinate transformation mode, so ONNX models using it failed to build. Add it to CoordTransformer as a length-independent, ROI-free coordinate map ((x + 0.5) / scale).

[kali]

/ci full

[github-actions]

🔴 Bench vs main — 1 speed regression(s) · ⚠️ 7 secondary

Reference: 2026-08-06 morning nightly run (0d old) · full report → run

Speed — evaltime · prefill · decode

Δ	metric	device	main → PR
🔴 +6.5%	arm_ml_kws_cnn_m evaltime · pass	cortex-a9	9.23 ms → 9.82 ms

⚠️ 7 secondary regression(s)

Δ	metric	device	main → PR
⚠️ +33.7%	arm_ml_kws_cnn_m load · pass	cortex-a9	83 ms → 111 ms
⚠️ +29.5%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a9	129 ms → 167 ms
⚠️ +13.9%	arm_ml_kws_cnn_m load · pass	cortex-a7	79 ms → 90 ms
⚠️ +11.8%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a7	119 ms → 133 ms
⚠️ +7.3%	hey_snips_v4_model17_nnef load · pulse8	cortex-a55	382 ms → 410 ms
⚠️ +7.0%	hey_snips_v4_model17_nnef load · pulse8	cortex-a53	471 ms → 504 ms
⚠️ +6.2%	en_tdnn_15M_nnef load · pulse_240ms	cortex-a53	452 ms → 480 ms

[github-actions]

❌ CI / full: failure

• ✅ old-harness: success
• ❌ cli-tests: failure
• ✅ onnx-tests: success
• ✅ tflite: success
• ✅ some-tests-with-paranoid-asserts: success
• ✅ without-default-features: success
• ✅ complexes: success
• ✅ check-all-targets: success
• ✅ C: success
• ✅ python: success

View workflow run

[kali]

⚠️⚠️⚠️ Just rebased! ⚠️⚠️⚠️

[kali]

/ci full

[czoli1976]

@kali nice, can you try this Metal Resize after this #2556 ?

[github-actions]

🔴 Bench vs main — 2 speed regression(s) · ⚠️ 12 secondary

Reference: 2026-08-06 morning nightly run (0d old) · full report → run

Speed — evaltime · prefill · decode

Δ	metric	device	main → PR
🔴 +10.1%	hey_snips_v31 evaltime · 400ms	cortex-a9	4.08 ms → 4.49 ms
🔴 +7.3%	hey_snips_v31 evaltime · 400ms	cortex-a7	5.3 ms → 5.69 ms

⚠️ 12 secondary regression(s)

Δ	metric	device	main → PR
⚠️ +20.5%	arm_ml_kws_cnn_m load · pass	cortex-a9	83 ms → 100 ms
⚠️ +17.4%	hey_snips_v4_model17 load · 2sec	cortex-a7	4.31 s → 5.06 s
⚠️ +15.4%	hey_snips_v4_model17 load+optimize · 2sec	cortex-a7	5.04 s → 5.81 s
⚠️ +15.0%	hey_snips_v31 load+optimize · 400ms	cortex-a7	381 ms → 438 ms
⚠️ +14.7%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a9	129 ms → 148 ms
⚠️ +14.7%	hey_snips_v31 load · 400ms	cortex-a7	334 ms → 383 ms
⚠️ +10.8%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a53	74 ms → 82 ms
⚠️ +10.4%	arm_ml_kws_cnn_m load · pass	cortex-a53	48 ms → 53 ms
⚠️ +8.1%	hey_snips_v31 load · 400ms	cortex-a9	298 ms → 322 ms
⚠️ +7.6%	hey_snips_v31 load+optimize · 400ms	cortex-a9	343 ms → 369 ms
⚠️ +6.9%	en_tdnn_15M_nnef load · pulse_240ms	cortex-a53	452 ms → 483 ms
⚠️ +6.8%	hey_snips_v4_model17_nnef load · pulse8	cortex-a53	471 ms → 503 ms

[kali]

/ci full

[github-actions]

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

Reference: 2026-08-06 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
⚠️ +28.9%	arm_ml_kws_cnn_m load · pass	cortex-a9	83 ms → 107 ms
⚠️ +21.7%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a9	129 ms → 157 ms
⚠️ +11.5%	hey_snips_v4_model17_nnef load · pulse8	cortex-a55	382 ms → 426 ms
⚠️ +7.0%	hey_snips_v4_model17_nnef load · pulse8	cortex-a53	471 ms → 504 ms
⚠️ +6.9%	en_tdnn_15M_nnef load · pulse_240ms	cortex-a53	452 ms → 483 ms