Use PytorchModelHubMixin to save models as safetensors

[alexander-soare]

What does this PR do?

Primary changes:

• Make all policies subclass PytorchModelHubMixin as well as nn.Module.
• Remove save and load from all policies.
• Update logger.py to use save_pretrained instead of save, and to separately save the Hydra config as a yaml file.
• Update eval.py to use from_pretrained (this consolidates the two user paths into one, at the expense of some conditional branching in the code).
• Update examples to conform.
• Update README section on uploading policies.
• Remove "pretrained_model_path" from yaml config.

Side changes:

• Add get_policy_and_config_classes to avoid repetition of the conditional branching across policy names.
• Change ACTConfig.d_model to ACTConfig.dim_model for consistency with ACTConfig.dim_feedforward.
• Change model saving directory from "models" to "checkpoints" to avoid confusion between "policy" and "model" and be more explicit.
• Add functionality to eval.py to be able to evaluate a non-pretrained model.

Not for this PR:

• TDMPC.
• Script for uploading training checkpoint to the hub.
• Model cards.

How was it tested?

• CI

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR. Try to avoid tagging more than 3 people.

[Cadene]

Left some comments. Nothing blocking merging. Nice work!!!

[Wauplin]

Hey there! Sorry I only spotted this PR now :) Nice work to use PytorchModelHubMixin to integrate with the Hub!

I left a comment below as I think it can be even more tightly integrated for better documentation/discoverability on the Hub. In general, never hesitate to ping me (@Wauplin) in any PR that is related to the Hub. I'm maintain huggingface_hub and I'd be happy to help if you have questions about how other HF libraries are doing (and stay consistent with them) 🤗