Add HF integration

[NielsRogge]

Hi @wenbowen123, congrats on this work!

This PR ensures the FoundationStereo model can be pushed and loaded to and from the 🤗 hub.

It equips the main model class with:

• from_pretrained, save_pretrainedandpush_to_hub capabilities
• safetensors for weights serialization (if you do model.save_pretrained("...") or model.push_to_hub("..."))
• track download numbers for models pushed to the hub, similar to models in the Transformers library
• push an automated model card.

It leverages the PyTorchModelHubMixin class which allows to inherits these methods.

Usage is as follows:

from core.foundation_stereo import FoundationStereo

model = FoundationStereo(...)

# save it locally like so
model.save_pretrained("my-foundation-stereo-model")

# or push a trained model to the hub
model.push_to_hub("nvidia/foundation-stereo")

# now anyone can use it like so
model = FoundationStereo.from_pretrained("nvidia/foundation-stereo")

[NielsRogge]

Disclaimer: please don't merge this PR before trying to push and reload to and from the hub to ensure it works as expected :) thank you