[Common] Enable arch support for Rubin

[denera]

Changes

• Gate arch-specific Rubin codegen on CUDA 13.4+.
• Drop redundant cute includes from the cast-fusion kernels that caused compile errors on cute::maybe_ifence() and cute::maybe_warp_switch().
• Enable stochastic rounding for ArchSpecific<107>.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Summary

The PR prepares the common CUDA library for Rubin.

• Adds SM107a architecture-specific compilation for CUDA 13.4 and newer.
• Enables stochastic rounding for architecture 107.
• Removes unused CUTLASS includes from Hadamard-transform fusion sources.

Confidence Score: 5/5

The PR appears safe to merge.

The corrected CUDA toolkit guard now reaches the SM107a target for default CUDA 13.4 builds, and no blocking failure remains.

Important Files Changed

Filename	Overview
transformer_engine/common/CMakeLists.txt	Adds a valid CUDA 13.4 version guard that appends 107a when architecture 100 family compilation is enabled; the previously reported malformed guard is fixed.
transformer_engine/common/util/ptx.cuh	Extends the existing architecture-specific stochastic-rounding capability check to architecture 107.
transformer_engine/common/hadamard_transform/graph_safe_group_row_cast_col_hadamard_transform_cast_fusion.cu	Removes two unused CUTE includes without changing kernel behavior.
transformer_engine/common/hadamard_transform/group_hadamard_transform_cast_fusion.cu	Removes two unused CUTE includes without changing kernel behavior.
transformer_engine/common/hadamard_transform/group_row_cast_col_hadamard_transform_cast_fusion.cu	Removes two unused CUTE includes without changing kernel behavior.
transformer_engine/common/hadamard_transform/hadamard_transform_cast_fusion.cu	Removes two unused CUTE includes without changing kernel behavior.
transformer_engine/common/hadamard_transform/row_cast_col_hadamard_transform_cast_fusion.cu	Removes two unused CUTE includes without changing kernel behavior.

_{Reviews (3): Last reviewed commit: "[Common] Enable arch support for Rubin" | Re-trigger Greptile}