[Common] Move arch specific helpers in ptx.cuh to a new header

[kainzhong]

Description

ptx.cuh current includes some helper that can only be compiled with arch-specific flags.
For example try_cancel_cta where there is constexpr bool is_blackwell = ARCH_BLACKWELL_FAMILY; which will instantiate ArchSpecific / FamilySpecific template and trigger the static assertion which will fail if you compile with generic arch flags like "sm100".
This will take effect as you include the header, even if you just use some other functions that are totally arch agnostic (static assertion happens when the template is instantiated regardless of whether you actually use it or not).

This PR moves the arch specific helpers to a new header so now if you need arch agnostic helpers, you should include ptx.cuh, and if you need arch specific helpers, you include ptx_arch_spec.cuh

Also, ptx.cuh is not a public header so this won't be a breaking change

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

Changes

• ptx.cuh is an arch agnostic header now that can be included by files in transformer_engine_cuda_sources, which are compiled with generic flags (NVTE_GENERIC_ARCHS)
• ptx_arch_spec.cuh is an arch agnostic header now and it can only be included by files in transformer_engine_cuda_arch_specific_sources, which are compiled with specific flags (NVTE_SPECIFIC_ARCHS)

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 separates architecture-specific PTX helpers from the architecture-agnostic utility header.

• Moves architecture- and family-specific helpers into ptx_arch_spec.cuh.
• Updates affected CUDA sources to include the architecture-specific header.
• Correctly documents that the new header is limited to transformer_engine_cuda_arch_specific_sources.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
transformer_engine/common/util/ptx.cuh	Removes architecture-specific helpers while retaining architecture-agnostic PTX utilities.
transformer_engine/common/util/ptx_arch_spec.cuh	Houses the relocated architecture-specific helpers and now gives correct source-group guidance.
transformer_engine/common/cast/mxfp8/quantize_mxfp8.cuh	Switches the MXFP8 quantization implementation to the architecture-specific PTX header.
transformer_engine/common/cast/nvfp4/quantize_transpose_nvfp4.cuh	Switches the NVFP4 quantization implementation to the architecture-specific PTX header.
transformer_engine/common/multi_tensor/compute_scale.cu	Uses the architecture-specific header required by its relocated helper dependency.
transformer_engine/common/recipe/nvfp4.cu	Uses the architecture-specific header required by NVFP4 recipe code.

_{Reviews (4): Last reviewed commit: "Update transformer_engine/common/util/pt..." | Re-trigger Greptile}

[timmoon10]

Overall LGTM

[ptrendx]

/te-ci