Add Ascend grouped matmul direct2560 bundle

[xieminghui187]

Summary

• add the complete Ascend grouped-matmul direct2560 operator bundle for reuse by DLInfer
• adapt the implementation from CANN ops-transformer commit 8fff8f9279086404073e2b38e51474792cae9e7b
• include the renamed ACLNN/op-api entry, pybind wrapper, tiling implementation, and Ascend 910_93 artifacts
• document the source and adaptation at the operator entry

Validation

• FP16 and BF16 correctness: max_abs=0.0
• BF16 repeated-call test: 100 iterations passed
• Python compilation, clang-format, and git diff --check passed
• DLInfer ShareGPT 500-request output throughput: 1025.56 tok/s -> 1161.19 tok/s (+13.22%)

This PR is intentionally a source/binary bundle matching the DLInfer copy; DLBlas build-system integration can be handled separately when the public operator API is finalized.

[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.}