[Nebulixeay]

This PR adds an initial Cambricon MLU provider under backends/mlu, using InfiniTrain's PrivateUse1 interface and the repository's provider build entry point. It includes runtime integration, BANG C kernels, a GPT-2 example adapter and small operator regression tools. The root dispatcher, MACA provider and pinned framework revision are unchanged.

Work in progress: full GPT-2 single-card training for 10 steps and the requested end-to-end precision targets have not been verified. GEMM still uses a CPU fallback; BF16 training and multi-card support remain unfinished.

Build and validation

• Validated on MLU590 with GCC 13.4.0, cncc 5.7.6, CMake 4.4.5 and Ninja 1.13.2.
• Fresh configuration and builds of mlu_gpt2 and mlu_op_test succeeded against the pinned InfiniTrain v0.7.0 content (c90338c); all nine BANG C sources compiled.
• The 35 small-shape operator comparisons, host comparator self-checks and device-math tests passed. Finite/NaN injection tests returned the expected failure code; the controlled fatal test aborted with the expected diagnostic. Linear tests validate the host fallback.
• gpt2 --help successfully printed usage. No model training step was run during this validation.
• The remote staging tree used an archive of the fixed framework tree and shallow retrieval of the four dependencies at their pinned commits. Review compared 2,687 tracked files with the corresponding Git objects: content matched after CRLF/LF normalization. This validates the staged build content, not the standard recursive-submodule download procedure.
• No MACA or CUDA build regression was performed. Leaving those provider files unchanged is not a claim of cross-platform regression coverage.

Current limitations

• Device GEMM integration is pending. Optional CNNL handle creation/queue binding is available, but current operators do not compute through CNNL.
• Several kernels use serial execution; large training shapes and full-model backward remain unverified.
• BF16 Embedding accumulation still needs correction; no BF16 accuracy result is claimed.
• CNCL/multi-card support and shared InfiniTrain MLU test-suite integration are pending. BUILD_TEST=ON is explicitly rejected; use the included MLU regression tools.
• Strict end-to-end loss acceptance tooling is deferred to a separate change.

Model/data assets, build products, internal review scripts and development logs are excluded from this PR. Details and commands are in backends/mlu/README.md and backends/mlu/docs/validation.md.