Clarify missing SM100 sync-object factory errors in custom CuTe pipelines

Summary

Update the CuTe DSL sync-object factory selection in tensorrt_llm/_torch/cute_dsl_kernels/blackwell/custom_pipeline.py.

The current fallback assigns PipelineAsync._make_sync_object when cutlass.pipeline.sm100.PipelineTmaUmma._make_sync_object is unavailable. That fallback does not support PipelineOp.TCGen05Mma. The affected custom pipelines can then fail with an assertion that does not identify the missing SM100 factory.

Required changes

• Do not fall back to PipelineAsync._make_sync_object for the custom UMMA pipelines.
• Raise a clear import-time or runtime error when cutlass.pipeline.sm100.PipelineTmaUmma or its _make_sync_object factory is unavailable.
• Ensure PipelineTmaUmma.create, PipelineUmmaAsync.create, and PipelineCpAsyncUmma.create report the dependency problem clearly.

Rationale

The current merge-back matches the validated source branch. Track the compatibility improvement separately with the kernel author.

Affected area

• tensorrt_llm/_torch/cute_dsl_kernels/blackwell/custom_pipeline.py

Acceptance criteria

• Unsupported CuTe DSL installations fail with an explicit and actionable error.
• Supported installations continue to use the SM100 sync-object factory.
• The custom pipelines do not select PipelineAsync._make_sync_object for PipelineOp.TCGen05Mma.

Backlinks

• Pull request: https://github.com/NVIDIA/TensorRT-LLM/pull/18369
• Review discussion: https://github.com/NVIDIA/TensorRT-LLM/pull/18369#discussion_r3881387655
• Requested by: @farazkh80