Switch to event based/callback based approach for dream web

[TesseractCat]

This uses HTTP fetch streaming to push images to the web interface as they are generated, and adds a progress bar by exposing the img_callback parameter to prompt2image.

As a result of using a callback based approach, when generating multiple images, it will add them to the result list immediately, rather than waiting until all images have completed.

[bakkot]

Here is a commit on top of this branch which also streams the intermediate images, so you can watch it denoise. Feel free to pull this in to this branch.

(This really makes me want to have the ability to cancel requests, because you can often see it's not going to be what you want before it actually finishes.)

[lstein]

I'm fixing issues with the dream_web having to do with upscaling and will work on this next.

[lstein]

Unfortunately in the intervening time between submission of the PR, dream_web.py has been made into a module and moved into ldm/dream/server.py. scripts/dream.py --web now invokes this module. I think this makes better sense than maintaining two different launch scripts that share identical arguments.

Would it be possible to redo the PR using the new structure?

[bakkot]

I took the liberty of doing that rebase in #180. Feel free to ignore it if you'd prefer to keep this PR as the source of truth, though.