Add PyTorch version compatibility documentation and validation

[Changomango0903]

Currently, torchtitan releases implicitly depend on the latest cutting-edge PyTorch versions without explicit documentation or runtime validation. This creates several issues:

1. Users experience cryptic errors when running torchtitan with older PyTorch versions that lack required APIs
2. No clear compatibility matrix exists to help users determine which PyTorch version they need
3. Silent breakage can occur when PyTorch APIs change between versions
4. Difficult to debug version-related issues without upfront validation

[tianyu-l]

Thanks for the feedback. Could you say more about your use case?

without explicit documentation or runtime validation.

Technically

• it is explicitly mentioned https://github.com/pytorch/torchtitan/blob/main/README.md?plain=1#L19
• our CI is running with latest pytorch nightly, so it is validated

No clear compatibility matrix exists to help users determine which PyTorch version they need

torchtitan nightly works with pytorch nightly; for torchtitan "stable release", we pin pytorch versions https://github.com/pytorch/torchtitan/blob/main/README.md?plain=1#L120

Users experience cryptic errors when running torchtitan with older PyTorch versions that lack required APIs

Curious when would you hit such errors -- would it be helpful if we always publish a torchtitan stable release that works with pytorch stable release? But note that for this you don't get to use the latest features in between two pytorch releases.