Rendering regression test

[hrydgard]

We should collect a large number of GE captures and have a way to automatically render and compare them to blessed images. Dolphin has something similar that runs in the cloud, but I primarily want this for all the mobile devices - they generally have the worst driver bugs and rendering issues. It would run locally on the device but use network sharing to download and run each GE capture so we don't have to manually copy over files.

How about this now?

[unknownbrackets]

I think what we could do is:

1. Choose a batch of representative GE frame dumps (we have a bunch reported now.)

2. Generate representative screenshots for them. I have a tool that does this for (most) using PSP rendering, which is arguably the best thing. There's an argument for using a specific device for "best" due to higher bit rate, etc.

3. Problem: setup some form of CI? Maybe llvmpipe or something? I think we need to validate on pull request, or else we'll break it and not notice which would erode the value of these. This would be separately useful anyway, but Actions and Travis and etc. don't have GPUs. Obviously this is just for one (or a couple) reference passes, as long as those pass we could rely on users for the rest.

4. Decide if we're maintaining the list in code (seems better? but then we can't ever remove/change and might version...) or remotely (i.e. a URL that lists them.)

5. Expose a button or similar that runs each test in sequence. Question for this: do we force off hack settings for it? Or warn/block running it when settings are dangeresque?

6. Optional bonus feature: measure a baseline rendering speed for each frame dump, and provide some device performance score based on how the device's times compare.

I think the biggest problem is CI. It's tempting to ignore it, but I think the feature will fall flat without it.

-[Unknown]