Improve HF integration

[NielsRogge]

Hi @dbolya, @berniebear and team!

Thanks for making the models available on the hub. This PR improves the 🤗 integration, by making the main class inherit from the PyTorchModelHubMixin class. This was also adopted by your colleagues for the following works:

• https://huggingface.co/facebook/VGGT-1B
• https://huggingface.co/facebook/locate-3d
• https://huggingface.co/facebook/cotracker3
• https://huggingface.co/facebook/sam2.1-hiera-large
• https://huggingface.co/facebook/hiera-tiny-224-hf
• and more.

As well as other popular models:

• https://huggingface.co/sesame/csm-1b
• https://huggingface.co/naver/DUSt3R_ViTLarge_BaseDecoder_512_dpt

Benefits:

• weights are serialized in safetensors format, instead of pickle
• download stats work
• the class automatically inhertis the from_pretrained, save_pretrained and push_to_hub methods.

You can push weights to the hub as follows (for example, for https://huggingface.co/facebook/PE-Core-L14-336):

from model import CLIP
from huggingface_hub import hf_hub_download
import torch

# load model
model = CLIP(**model_cfg)

# equip with weights
filepath = hf_hub_download(repo_id="", filename="")
state_dict = torch.load(filepath, map_location="cpu")
model.load_state_dict(state_dict)

# push to the hub
model.push_to_hub("facebook/PE-Core-L14-336")

# now anyone can use it like so:
model = CLIP.from_pretrained("facebook/PE-Core-L14-336")

It would be great if you could try this out, and overwrite the weights of the model repository with the updated ones.