transformers: bound an attention mask by its known And factor

[kali]

A band mask ANDed with a padding mask built from a runtime length has no uniform_tdim, so ScaledMaskedSoftmax planted no ROI on the scores and the rel-pos attention could not be blockified for pulsing. input_roi now uses an upper bound of the mask: its uniform_tdim, or the known factors of an And, followed through axis ops and casts.

[github-actions]

⚠️ Bench vs main — no speed regressions · 13 secondary regression(s)

Reference: 2026-10-13 morning nightly run (0d old) · full report → run

Speed — evaltime · prefill · decode

no inference-speed regressions

⚠️ 13 secondary regression(s)

Δ	metric	device	main → PR
⚠️ +25.9%	nemotron_3_5_asr_streaming_0_6b_f32f32_encoder_1s load · cpu	apple-m1-max	3.44 s → 4.32 s
⚠️ +25.5%	nemotron_3_5_asr_streaming_0_6b_f32f32_encoder_1s load · metal	apple-m1-max	3.44 s → 4.32 s
⚠️ +25.0%	nemotron_3_5_asr_streaming_0_6b_f32f32_encoder_1s load · cpu_mt	apple-m1-max	3.45 s → 4.32 s
⚠️ +24.0%	nemotron_3_5_asr_streaming_0_6b_f32f32_encoder_1s load+optimize · metal	apple-m1-max	3.63 s → 4.5 s
⚠️ +23.1%	nemotron_3_5_asr_streaming_0_6b_f32f32_encoder_1s load+optimize · cpu	apple-m1-max	3.87 s → 4.76 s
⚠️ +23.0%	nemotron_3_5_asr_streaming_0_6b_f32f32_encoder_1s load · cuda	i9-11900kb_rtx-4060	4.73 s → 5.82 s
⚠️ +22.2%	nemotron_3_5_asr_streaming_0_6b_f32f32_encoder_1s load+optimize · cpu_mt	apple-m1-max	3.89 s → 4.75 s
⚠️ +19.9%	nemotron_3_5_asr_streaming_0_6b_f32f32_encoder_1s load+optimize · cuda	i9-11900kb_rtx-4060	5.47 s → 6.57 s
⚠️ +11.1%	hey_snips_v1 load · 400ms	cortex-a7	63 ms → 70 ms
⚠️ +6.0%	en_tdnn_lstm_bn_q7 RSS @ load · pulse_240ms	cortex-a53	25.1 MB → 26.7 MB
⚠️ +5.3%	en_tdnn_8M RSS @ ready · 2600ms	cortex-a55	63.8 MB → 67.2 MB
⚠️ +5.3%	en_tdnn_8M load · pulse_180ms	cortex-a55	982 ms → 1.03 s
⚠️ +5.3%	en_tdnn_8M load+optimize · pulse_180ms	cortex-a55	1.12 s → 1.18 s

---

Bench numbers come from shared CI hardware and do produce false positives: a flagged
row is not on its own evidence of a real regression, and a clean report is not proof
there is none. Contributors do not need to chase what is reported here — leave the
reading of it to the maintainers, who will say if something needs action.