[Main][Perf] Add fused matmul/reduce-scatter kernel for performance optimization.

[ZYang6263]

What this PR does / why we need it?

This PR boosts performance by introducing a fused kernel for the matrix matmul and reduce scatter operations. It supports both unquantized (e.g., BFloat16) and W8A8 quantized models.

Does this PR introduce any user-facing change?

How was this patch tested?

• vLLM version: v0.11.0rc3
• vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

[gemini-code-assist]

Code Review

This pull request introduces a fused matmul/reduce-scatter kernel to optimize performance, which is a great initiative. The implementation correctly identifies the specific CANN version and quantization methods to apply the optimization. However, I've found a critical correctness issue where the bias term is ignored in the unquantized path, which could lead to incorrect model outputs. Additionally, there are several opportunities to improve code quality by removing dead code, placing imports correctly, and refactoring duplicated logic. Addressing these points will make the new optimized code paths more robust and maintainable.

[gemini-code-assist]

The output_parallel tensor initialized here is never used, as the result from npu_mm_reduce_scatter_base is assigned directly to the output variable. This is dead code and should be removed. A similar redundant initialization exists for the W8A8 path on lines 432-436.

[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.