feat: support Ascend950 attention and paged KV cache.

[wuyifan2017]

Description

• Route Ascend950 paged KV cache writes through the ReshapeAndCacheA5 wrapper and register the wrapper in the NPU kernel target.
• Add Ascend950 PyTorch SDPA fallbacks for packed TND prefill and paged decode attention, including GQA KV-head expansion.
• Update the xllm-ops submodule to the implementation from xLLM-AI/xllm-ops.
• Correct the documented paged-cache layout and add Qwen3.5-27B TP4 coverage for full attention and a KV cache crossing the 128-token block boundary.

Related Issues

Depends on xLLM-AI/xllm-ops#31.

Change Type

• Bug fix
• New feature
• Performance improvement
• Refactor
• Documentation
• Test
• Build or CI

Pull Request Checklist

PR Title and Commit Messages

• The PR title and each commit message follow the xLLM commit format: <type>: <subject>.

Pre-commit Checks

• I have installed pre-commit by running pip install pre-commit or an equivalent command.
• I have installed the hooks with pre-commit install.
• I have run pre-commit run --all-files and fixed any reported issues.

Self Review

• I have self-reviewed the code according to .agents/skills/code-review/references/custom-code-style.md, especially code written or assisted by AI.
• I have rebased this PR onto the latest main branch.

Build and Test Coverage

• Tests have been added or updated as needed.
• CUDA: python setup.py build test has passed on a CUDA machine.
• NPU: python setup.py build test has passed on an NPU machine.
• MLU: python setup.py build test has passed on an MLU machine.

Reviewer Notes

• The fallback covers non-chunked packed TND prefill and standard paged decode on Ascend950. ACL graph decode remains unchanged.
• pre-commit run --all-files passes.

[yingxudeng]

is_ascend950() 和 expand_kv_heads 不同文件完全重复定义。可以考虑放到 core/kernels/npu/utils.h 或者其他合适位置

[wuyifan2017]

已修改