Switch to event based/callback based approach for dream web (rebased)

[bakkot]

This is #129 rebased, plus my commit which adds renders images as they denoise (which is fun, and also lets you see if the image is going to be entirely the wrong shape, though as of now it's impossible to cancel a generation so this is not helpful) (EDIT: my commit has been moved to #181).

[TesseractCat]

This rebase looks good, but the progress visualization needs to be optional for performance reasons.

If there's a way to asynchronously read tensor data from the GPU we should do that.

[lstein]

This looks good and will be generally useful outside the webUI as well. Any idea what the performance hit is?

[lstein]

I'm almost done with the integration. There were some issues with correctly handling upscale requests, which should replace the original rather than supplement it. (can add option to keep originals later)

[bakkot]

Any idea what the performance hit is?

No noticeable performance hit from this PR. I pulled out the slow part to #181, which is about a 20% slowdown when the checkbox is checked (which it is not by default).

[lstein]

This turned out to be harder than expected due to the way in which a batch of images are first generated, and then post-processed using GFPGAN and ESRGAN. However, it seems to be working now. When post-processing is requested, the code first generates and displays the original image, then displays a "post-processing file 1/3" message, and then replaces the original image with the upscaled image. The only problem I've encountered is that the browser seems to cache the image thumbnail, and may not display the reconstructed faces when the underlying image files are updated. This will become a non-issue when PR #181 is integrated.

Speaking of which, there are likely to be new conflicts with PR #181, and I apologize in advance if I've created a headache for @bakkot .

[lstein]

This was inadvertently closed before the merge was pushed. However, the PR is now merged into main.