feat(cambricon): integrate InfiniOps backend

[baominghelly]

Summary

• enable the InfiniOps runtime, context, and dispatcher for Cambricon
• initialize RoPE cache tensors for the Cambricon InfiniOps path and register the rotary embedding adapter
• use InfiniOps Argmax for Cambricon greedy sampling and synchronize Cambricon D2H copies before host reads
• extend the InfiniOps external build for Cambricon, including the selected INFINI_OPS_OPS adapters, PyTorch/torch_mlu ABI and link settings, and a single dispatch translation unit

Verification

Tested in infinicore-cambricon-dev-libaoming with:

export INFINI_OPS_OPS=add,argmax,copy,embedding,fused_add_rms_norm,gemm,rms_norm,rotary_embedding,silu_and_mul
xmake build
xmake install
xmake build _infinicore
xmake install _infinicore

The complete build/install chain exited successfully. Both imports below succeeded with the generated extension:

python3.10 -c "import infinicore; print(infinicore.__file__)"
python3.10 -c "import torch; import torch_mlu; import infinicore; from infinicore.lib import _infinicore; print(_infinicore.__file__)"

Related InfiniOps operator PRs:

• Copy: feat(cambricon): add copy provider InfiniOps#927
• Rotary Embedding: feat(cambricon): add rotary_embedding provider InfiniOps#944

[wooway777]

conflict了，另外xmake.lua额外需要这么多修改的理由是什么？有些东西是不是应该收束到bang.lua里面

[baominghelly]

conflict了，另外xmake.lua额外需要这么多修改的理由是什么？有些东西是不是应该收束到bang.lua里面

conflict已解决。xmake.lua之前有一些清理之前编译缓存的操作，现在已删除；剩下的还有一个算子白名单的功能，因为寒武纪目前算子偏少，不加的话导入会报错算子找不到，这个功能等下一阶段算子补充好并且模型验证完后会去掉。