Preserve Hugging Face metadata when checkpointing

[dylantirandaz]

Summary

• Download text model files so tokenizer merges are available during checkpoint initialization
• Preserve tokenizer, config, and custom Python files when writing checkpoints
• Add focused coverage for model download selection and checkpoint metadata filtering

Fixes #631

Testing

• cargo check -p psyche-client -p psyche-data-provider --tests --features tch/doc-only
• git diff --check