Ensure weight transpose is valid for Hopper FP8 training

[guyueh1]

Description

Currently for Hopper FP8 training, when weight is Float8Tensor and weight._transpose is invalidated by the optimizer step, no mechanism handles the creation of weight transpose thus the Dgrad errors out. This PR updates weight usage and ensure columnwise data is there before saved for bprop.

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

[guyueh1]

@ksivaman @ptrendx please review, this is to fix the hopper fp8 error: at the beginning of a new step, the weight transpose cache is invalid so the backprop errors out. I am specifiying weight to have column-wise usage to make sure it creates transpose if it's on hopper.

[timmoon10]

LGTM

[timmoon10]

/te-ci pytorch L1

[guyueh1]

@timmoon10 I looked at one of the failed L0 test (unittest H100 1GPU) and it's failing for a module I didn't change (page attention). Is L0 ut expected to pass?