hotfix - Revert vllm/attention/layer.py changes from 0f8cafe - fix torch.compile recompilations

hotfix - Revert vllm/attention/layer.py changes from 0f8cafe - fix torch.compile recompilations #1571