Custom API endpoints

[bfirsh]

It'd be neat if you could define custom API endpoints that could be accessed via the HTTP server or via the queue server.

Maybe it's as simple as documenting how to create a custom HTTP server that extends the FastAPI server, which is totally possible right now without any changes to Cog.

User data

• @postera-james wants to use this as a way to fine-tune a model live. They specifically don't want to save a whole Docker image each time because that's too heavyweight. When the model is updated, they could then use docker commit to save the model. A stop gap for #647 (which might not be the right solution anyway because it implies saving a whole new Docker image each time).

[massile]

@bfirsh Is it still possible to extend the FastAPI endpoints? Looks like the app object in the http file is not accessible from outside.

https://github.com/replicate/cog/blob/main/python/cog/server/http.py#L505