Implement RAVU / FSRCNN / NNEDI3 / SuperXBR / whatever

Too many memes, too little time. Most of these are open source and have more-or-less compatible licenses, so we could potentially attempt implementing them natively. cf. https://github.com/bjin/mpv-prescalers

Might be scrapped in favor of letting users provide them manually, see #19 (closed). An open question would be figuring out how we want to handle the trained weights - one idea would be to keep them in a separate repository as a git remote and make it an optional build-time dependency.