[cuda backend] skip fully-masked KV blocks calculation in SDPA#20198

Gasoonjia merged 8 commits into

mainpytorch/executorch:mainfrom

g4-opt-prefill-window-sdpapytorch/executorch:g4-opt-prefill-window-sdpaCopy head branch name to clipboard