[Core] Depcrecate torch_dtype from from_pretrained

[sayakpaul]

What does this PR do?

We recently fully deprecated (#6886) torch_dtype and torch_device when using to(). But we still have torch_dtype in from_pretrained(). This is a point of confusion for our users. This PR, hence, starts a depcrecation cycle for torch_dtype from from_pretrained().

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[patrickvonplaten]

Generally I like this change, but IMO we should make sure that we're aligned with Transformers here and this is currently not the case: https://github.com/huggingface/transformers/blob/de11e654c962d5b23eb53a4387cd637b01987491/src/transformers/configuration_utils.py#L273

Also see: https://huggingface.slack.com/archives/C01N44FJDHT/p1701688646192709 (I've never really followed up here)

[sayakpaul]

Agree that both the libraries should have similar modeling behaviour. However, this particular case could lead to a confusing developer experience IMO. In one place, we have dtype (to()) and in another (from_pretrained()) we have torch_dtype.