Support local paths for VAE and prompt encoder URIs

[yondonfu]

Context

For my use case, I want to specify a local path instead of relying on the default HF hub cache at ~/.cache/huggingface/hub.

Summary

• Add ae_uri and prompt_encoder_uri parameters to WorldEngine constructor for flexible model loading
• Fallback to local path in InferenceAE.from_pretrained when HuggingFace download fails
• Allow overriding default config URIs at runtime

Test plan

• Verify loading models from HuggingFace Hub still works
• Verify loading models from local paths works

🤖 Generated with Claude Code

[lapp0]

Can we make these part of self.model_cfg, which can be updated via model_config_overrides?

The ae and prompt encoder are bound to the model.

[yondonfu]

This is fixed and now the local paths can be passed with this syntax:

WorldEngine(
    model_config_overrides={
        ae_uri=...
        prompt_encoder_uri=...
    }
)

Example