Add support for autocast custom ops in GaudiTrainer

[regisss]

What does this PR do?

For some models, specifying custom bf16 and fp32 ops for Torch Autocast can lead to significant speedups. This PR enables users to specify such custom ops the same way as with HMP and they will be recognized by the trainer.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[ZhaiFeiyue]

@regisss LGTM! 👍
one question, when will you decided to remove HMP totally, or just leave it there?

[regisss]

@ZhaiFeiyue The only remaining obstacle is to make Wav2Vec2 training work. @libinta is taking a look at it in HS-1107.
Once this is cleared, I'll publish new Gaudi configs for GPT2 and Wav2Vec2 and release a patch to make sure everything works fine with autocast.
Finally, when SynapseAI 1.12 is released, I'll merge #349 to completely remove HMP.

[ZhaiFeiyue]

that's great!