Allow users to easily set IP-based rate limits

[abidlabs]

It would be nice if a user could set a flag to only allow prediction requests to come from the same domain as where the web server is running. In particular, if a user is hosting their model on Spaces, they may only want requests to come from other Spaces.

Internal discussion: https://huggingface.slack.com/archives/C02136Y252P/p1666630955157749?thread_ts=1666447456.974249&cid=C02136Y252P

cc @freddyaboulton @aliabid94 @osanseviero @radames @apolinario

[abidlabs]

We could consider re-working the api_open parameter to do this in 4.0

See https://huggingface.slack.com/archives/C02QZLG8GMN/p1695985205587149 (internal)

[abidlabs]

There are two ways we could implement this. Any preferences?

(1) Look at the referer header and ensure its the same as the server URL. This would not prevent Spaces to be called from another Space. These headers can also be spoofed pretty easily. (so doesn't completely prevent API usage).

(2) A more robust implementation would be to send a token from the backend to the frontend when the Gradio app is loaded. Making a prediction requires the frontend to send this token along with the prediction. These tokens have an expiry date (maybe 24 hours), so that someone can't simply read the token and hardcode it in the downstream application.

cc @osanseviero @apolinario @Wauplin @pngwn @freddyaboulton

[abidlabs]

Won't have a chance to do this for 4.0, but we can do this after 4.0 by adding another parameter. Moving to the 4.x milestone

[abidlabs]

I'm inclined towards changing this issue to instead allow developers to set ip-based rate limits as those are less spoofable. As that would be a separate parameter, we can remove from 5.0 milestone