Bundle direct2560 grouped matmul for Ascend

[xieminghui187]

Summary

• bundle the Ascend grouped-matmul direct2560 operator copied from the DLBlas implementation
• expose the ACLNN operator through pybind on the current NPU stream
• register the bundled, renamed custom OPP before importing torch_npu to avoid global operator collisions
• add a contiguous 2560-expert execution path that reduces each MoE MLP from six chunked grouped matmuls to two grouped matmuls
• retain the existing chunked implementation as the default; enable this path with DLINFER_GMM_EXPERIMENT=direct2560

The branch is based directly on the requested DLInfer main commit 92cd7982ccb142a18278be6692a6142740099631.

Implementation notes

• supports FP16/BF16, 2560 logical experts, and type-1 tensor group lists
• uses a renamed DlinferGroupedMatmulDirect ACLNN/op-api entry and bundled Ascend 910_93 artifacts
• fixes repeated-call teardown by respecting aclTensorList ownership and releasing the temporary unique_ptr guards
• includes a direct probe and unit coverage

Validation

• FP16 correctness: max_abs=0.0
• BF16 correctness: max_abs=0.0
• BF16 repeated-call test: 100 iterations passed
• TP=2 LMDeploy smoke test passed
• ShareGPT benchmark: 500/500 requests succeeded
• output throughput: 1025.56 tok/s -> 1161.19 tok/s (+13.22%)
• Black, clang-format, Python compilation, and git diff --check passed

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}