[MetaXGPU] Add compiler-path C500 hgemm route

[VitalyAnkh]

Hi maintainers,

This PR routes the C500 hgemm path in tileops.ops.gemm.GemmOp through the TileLang compiler-generated backend. The handwritten maca_hgemm MACA C implementation remains a performance, layout, and shape reference; it is not used as the optimized execution path from GemmOp.

Summary

• Adds the compiler-path MetaX C500 hgemm route with packed-B and split-K support.
• Keeps the helper and wrapper pieces needed to preserve the validated layout contract.
• Updates auto-dispatch coverage for the new GemmOp path.
• Requires TileLang 0.1.10, which provides the GEMM annotation surface used by this route.
• Keeps the route paired with the TileLang lowering and layout support in the companion PR.

Rebase update

• Rebased onto the latest upstream dev.
• Updated the direct hgemm block scopes to the current TileLang T.sblock API.
• Revalidated the paired TileOps/TileLang stack after the TileLang WSM contract fix.

Validation

• Production-shape hgemm sweep passed correctness on all 8 covered shapes.
• Current measured throughput range: 172.076209 to 205.316929 TFLOPS.
• Minimum A100-relative ratio in the sweep: 89.68%.
• A representative compiler-path WSM fallback guard passed correctness with the paired TileLang PR.

Notes

• Companion TileLang PR: [MetaxGPU][feature] Add MACA GEMM compiler-path layout support tile-ai/tilelang-metax#90
• The paired TileLang lowering and layout support is required for this route; both sides were validated together.
• The optimization stays on the compiler-generated MACA C implementation path rather than calling the handwritten reference kernel directly.

[gemini-code-assist]

Code Review

This pull request adds comprehensive support for MetaX C500 GPUs in GEMM operations, introducing specialized kernel paths such as BSM, split-K, and packed-B tile layouts. It includes several new kernel implementations, hardware-specific C++ headers, and auto-dispatch logic. The review feedback recommends defaulting compilation flags to enable MetaX-specific optimizations automatically on supported hardware and moving split-K validation logic to the initialization phase to prevent runtime crashes during tensor preparation.

[VitalyAnkh]

Addressed in the current head.

• _gemm_compile_flags now takes use_maca: Optional[bool] = None and defaults to runtime C500 detection when no explicit value is provided.
• The TILEOPS_GEMM_SPLIT_K compatibility checks for block_k now run immediately after init_config, before the prepared-B path can hit a less helpful shape error.

I will continue monitoring this PR together with the companion TileLang PR.