[bug-fix] Delete .pt checkpoints past keep-checkpoints

[ervteng]

Proposed change(s)

Update the checkpoint manager to manage non-onnx files as well. This means that PyTorch checkpoints also show up in the training_status.json, and are deleted when they expire.

Types of change(s)

• Bug fix
• New feature
• Code refactor
• Breaking change
• Documentation update
• Other (please describe)

Checklist

• Added tests that prove my fix is effective or that my feature works
• Updated the changelog (if applicable)
• Updated the documentation (if applicable)
• Updated the migration guide (if applicable)

Other comments

[vincentpierre]

Looks good except for the training status version update that I do not yet understand.

[vincentpierre]

Why bump this? It does not seem to be used anywhere. If this is needed, you need to check for backwards compatibility at some point no?

[ervteng]

I think we should still tick it. We don't check it b/c there hasn't been any breaking changes with the loading of this file - if we do make a breaking change, I would think we'll have to tick to 1.0 and check. Checking if the minor versions are the same was somewhat subsumed by the mlagents version check, so we didn't add a separate check for this file. At the end of the day this versioning was intended so that Cloud could reliably use this file if needed.

With that said I did tick it incorrectly - it should be 0.3.0 not a patch ver.

[vincentpierre]

Ok, feel free to merge when the tests pass