Add modal launch hf-download utility

[mwaskom]

Adds a new modal launch utility intended to make it simple to cache an asset from the Hugging Face hub onto a modal.Volume.

Backward/forward compatibility checks

---

Check these boxes or delete any item (or this section) if not relevant for this PR.

• Client+Server: this change is compatible with old servers
• Client forward compatibility: this change ensures client can accept data intended for later versions of itself

Note on protobuf: protobuf message changes in one place may have impact to
multiple entities (client, server, worker, database). See points above.

---

Changelog

• Added a modal launch hf-download utility. This provides a simple CLI for caching an asset from the Hugging Face Hub onto a Modal Volume:

  modal launch hf-download stabilityai/stable-diffusion-xl-base-1.0 hf-hub-cache --secret=hugging-face
  

  To use the assets, mount the Volume to a location that corresponds to the Hugging Face cache directory, as specified by the HF_HUB_CACHE environment variable or passed as the cache_dir in relevant functions.

[erikbern]

I'm not entirely sold on this as a utility vs just making sure our examples show how to download stuff. Feels like this as a tool hides the more general principle of how to run data prep in Modal. Also with HF, you can rely on lazy caching in most cases.

[mwaskom]

I agree that we should make it clear that Modal can be used throughout the project lifecycle! But this step doesn't seem like a very interesting "preprocessing" step to demonstrate; it's something you'd need to put in every App and it's pure boilerplate.

Relying on implicit caching can also work, but there's a couple advantages here:

• We're enabling the rust-based hf-transfer download, which is a lot faster, but requires a few extra steps that aren't otherwise relevant for Apps that use the models
• Writing a Model class so that you can trigger a download as a side effect of spinning up a container isn't always trivial; you need to write some nullary endpoint (which might not be the natural way to write your inference interface), and the download will happen in a container that requests resources you need to actually serve the model, which might be overkill (e.g. it's wasteful to have idling GPUs while the download is happening)