[PyTorch][Mcore] Fix illegal memory access issue while using Mcore async checkpoint with fp8 tensorwise recipe

[zhongbozhu]

Description

Replace the tex.dequantize with a torch implementation fixes the issue for now. Need to root cause why tex.dequantize did not work for the async cp, but working for the sync cp.

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

[timmoon10]

We should understand the root cause of why the FP8 dequantize is failing.

[zhongbozhu]

Update:

tex.dequantize not working for Mcore's async cp is likely because the kernel is launched in a spawned process instead of a forked process. Calling a torch implementation can get away with it since pytorch internally might handle this case well.

We can either merge this PR but it's likely that we will do the same thing for the dequantize kernel of mxfp8 recipe. Or we can root cause why the kernel launch doesn't work in the spawned process and fix this issue once for all.