Support model downloading/caching

[carlosgjs]

The models are too big to ship with the package, so a better approach is to download them on first use and cache them locally.

Subtasks:

• Create an org in Huggingface
• Upload the models to the org (two versions? 7b and 7b quantized?
• Create a mapping of a model version+platform to physical model file
• Add logic to download on-demand
• Support model versioning and checking for newer models