[Refactor] Remove redundant attention operator branches.

[weijinqian0]

[Refactor] Remove redundant attention operator branches.

Reason：

1. We replace other attention ops with fused_infer_attention_score expect decode_only state.
2. clean code and remove 310P support.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

This pull request refactors the attention operator branches by unifying prefill logic into a single _forward_prefill method and removing support for 310P devices. The changes simplify the control flow in the main forward method, making the code cleaner and easier to maintain. My review focuses on the correctness and clarity of this refactoring. I've identified one high-severity issue related to an unused parameter in the new _forward_prefill method, which should be addressed to improve code quality.

[gemini-code-assist]

The output parameter in the _forward_prefill method is unused. The value passed to it is immediately shadowed by the assignment output, _ = torch_npu.npu_fused_infer_attention_score(...) on line 357. This is misleading as it suggests an in-place operation which is not happening, and the pre-allocated tensor is wasted.

To avoid confusion and make the code cleaner, the output parameter should be removed from the function signature. The call site at line 592 should also be updated to no longer pass this parameter.

    def _forward_prefill(self,
                             query: torch.Tensor,
                             key: torch.Tensor,
                             value: torch.Tensor,
                             kv_cache: Tuple[torch.Tensor],
                             attn_metadata: AscendMetadata,
                             num_tokens=0):

[wangxiyuan]

replace by #4524