SkyPilot adds Hugging Face Storage as a zero-egress backend

SkyPilot has integrated Hugging Face Storage as a first-class backend, allowing AI workloads to run on any cloud while keeping models and datasets on the Hub with zero egress fees. This integration enables users to mount Hugging Face Buckets or repositories into SkyPilot jobs using a single `hf://` URL scheme.

• Mounts read-write Hugging Face Buckets or read-only model/dataset repos via `MOUNT` or `COPY` modes.
• Runs jobs across 20+ clouds, Kubernetes, Slurm, and on-prem clusters by leveraging SkyPilot's multi-cloud scheduling.
• Eliminates egress costs for reading data, as Hugging Face Storage charges no egress or CDN fees.
• Uses Xet-backed deduplication to store only changed chunks for incremental checkpoints and model variants.

This allows teams to utilize reserved or on-demand GPUs across multiple vendors without being pinned to a specific cloud's storage region, reducing costs and complexity.