Only call torch.autocast if it will have an effect

[hmellor]

Adds a context manager to replace direct use of torch.autocast that only autocasts if:

• autocast is already enabled in this context
• Or this call to maybe_autocast has enabled=True

This prevents autocast being added to the graph when it is effectively a no-op.
Which makes graph splitting in torch.compile more flexible as it removes the requirement that partition IDs be monotonically increasing.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: afmoe, apertus, arcee, aria, bamba, bitnet, blt, chameleon, cohere, cohere2, csm, cwm, dbrx, decision_transformer, deepseek_v2, deepseek_v3

[ArthurZucker]

Perfect thanks for pushing this!