[Bugfix] Zero-init MLA attention output buffers to prevent NaN from CUDA graph padding#37442

tlrmchlsmth merged 4 commits intovllm-project:mainvllm-project/vllm:mainfrom

elvircrn:zero-init-attn-output-prelvircrn/vllm:zero-init-attn-output-prCopy head branch name to clipboard