Fix scaled_mm CuTe compile contract for blockscaled persistent GEMM

[coderabbitai]

Summary

scaled_mm in tensorrt_llm/_torch/cute_dsl_kernels/blackwell/dense_blockscaled_gemm_persistent.py compiles Sm100BlockScaledPersistentDenseGemmKernel directly. Its positional arguments do not match Sm100BlockScaledPersistentDenseGemmKernel.__call__.

Required change

Update scaled_mm to compile the wrapper interface. Supply the required runtime dimensions and alpha_tensor. Preserve the existing pointer construction, layout mapping, cluster limit, stream, epilogue operation, and compilation options.

Rationale

The direct cute.compile(gemm_obj, ...) call binds the layout tuple where __call__ expects alpha. It then shifts the remaining positional arguments. This can produce an invalid compiled invocation.

Affected area

• tensorrt_llm/_torch/cute_dsl_kernels/blackwell/dense_blockscaled_gemm_persistent.py
• scaled_mm
• Sm100BlockScaledPersistentDenseGemmKernel.wrapper

Acceptance criteria

• scaled_mm compiles the wrapper with arguments that match the wrapper signature.
• The compile-time placeholders include the GEMM dimensions and alpha_tensor.
• Existing layout, pointer-alignment, stream, epilogue, cluster, and options behavior remains intact.
• Add or update focused coverage for the compile interface if the test environment supports it.

Backlinks

• Pull request: [None][feat] Add Rubin SM107 CuTe DSL foundation and BF16 kernels #18311
• Review comment: [None][feat] Add Rubin SM107 CuTe DSL foundation and BF16 kernels #18311 (comment)

Requested by: @farazkh80