[cuda backend] skip fully-masked KV blocks calculation in SDPA #7080

Job	Run time
CI run decision / decide	32s
Get changed files / get-changed-files	2s
export-model-cuda-windows-artifact (nvidia, parakeet-tdt, quantized-int4-weight-only) / linux-job	26m 57s
export-model-cuda-windows-artifact (mistralai, Voxtral-Mini-3B-2507, non-quantized) / linux-job	29m 47s
export-model-cuda-windows-artifact (facebook, dinov2-small-imagenet1k-1-layer, non-quantized) / linux-job	16m 26s
export-model-cuda-windows-artifact (mistralai, Voxtral-Mini-3B-2507, quantized-int4-weight-only) / linux-job	29m 54s
export-model-cuda-windows-artifact (mistralai, Voxtral-Mini-4B-Realtime-2602, quantized-int4-tile... / linux-job	30m 7s
export-model-cuda-windows-artifact (nvidia, parakeet-tdt, non-quantized) / linux-job	26m 3s
test-model-cuda-windows-e2e (mistralai, Voxtral-Mini-3B-2507, quantized-int4-weight-only) / windows-job	34m 41s
test-model-cuda-windows-e2e (facebook, dinov2-small-imagenet1k-1-layer, non-quantized) / windows-job	34m 2s
test-model-cuda-windows-e2e (nvidia, parakeet-tdt, quantized-int4-weight-only) / windows-job	34m 11s
test-model-cuda-windows-e2e (mistralai, Voxtral-Mini-4B-Realtime-2602, quantized-int4-tile-packed) / windows-job	35m 9s
test-model-cuda-windows-e2e (mistralai, Voxtral-Mini-3B-2507, non-quantized) / windows-job	35m 53s
test-model-cuda-windows-e2e (nvidia, parakeet-tdt, non-quantized) / windows-job	34m 39s
	6h 8m 23s

Provide feedback