Enable Xet for checkpoint model downloads

[luke-lombardi]

Summary

• stop forcing Hugging Face Hub onto the deprecated serial HTTP transfer path for checkpoint-enabled workloads
• keep Xet transfer scratch data on local ephemeral storage while model files remain in workspace storage
• cover transport environment behavior in gateway tests

Staging validation

• exact Qwen2.5-1.5B-Instruct cold deploy on RTX5090
• model transfer: 2.88 GiB in 22.3s
• untouched checkpoint archive: 7.97 GiB
• scale-to-zero restore and exact chat completion: HTTP 200 in 8.64s
• gVisor runtime restore: 6.74s

Tests

• go test ./pkg/gateway/services -count=1

---

Summary by cubic

Enable Xet for checkpoint model downloads to improve transfer performance and keep workspace volumes small. Move Xet transfer cache to ephemeral storage and let the runtime choose the Hub transport.

• New Features
  • Enable Xet for checkpoint-enabled model transfers; remove HF_HUB_DISABLE_XET and HF_HUB_ENABLE_HF_TRANSFER overrides.
  • Add HF_XET_CACHE in the ephemeral compile cache; keep model files in the workspace volume.
  • Extend gateway tests to cover env layout and transport selection.

^{Written for commit 35f34d4. Summary will update on new commits.}