Add NVTX ranges to categorize execution

[minitu]

Description

Adds NVTX ranges to categorize different parts of the execution.

Fixes # (issue)

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

Please list the changes introduced in this PR:

• Change A
• Change B

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

[minitu]

Please review @timmoon10

[timmoon10]

LGTM. I have some very minor renaming suggestions. We can merge once we fix merge conflicts and the TE main branch has been reconciled with release_v2.0.

For future reference, my comments on an earlier version of this PR: The overall design is delicate, but I can't think of a better approach. The torch.cuda.nvtx.range context is nicer than range_push/range_pop, but its CPU overhead is too high (1.7 us compared to 0.5 us).

[minitu]

@timmoon10 Sounds good, thanks!

[timmoon10]

I've added the option in nvtx_range_pop to explicitly specify the NVTX range, and it will throw an error if it gets something unexpected. This should help catch cases mismatches and mangled NVTX ranges. In the future, TE should replace its existing NVTX markers with these utility functions.

[timmoon10]

/te-ci pytorch