[REQUEST] Add Pytorch + CUDA version compatibility instructions

[eristocrates]

Like a few others I was also experiencing the issue of not seeing gpu usage despite doing the steps necessary to install pytorch. At first, I thought it was just the hardcoded cpu line but changing that to cuda actually gave me a NotImplementedError: Could not run 'torchvision::nms' with arguments from the 'CUDA' backend.

After a bit of old fashioned root cause analysis, turns out it's essential to get a compatible version of pytorch for your cuda version. I had cuda 12.5 and my pytorch was incompatible with it the whole time. This explains why torch shows up as available (cause it's installed) but doesn't actually get used even when mokuro displays cuda has been selected in the log. After i installed the nightly pytorch version for cuda 12.4(Note that you cannot install a PyTorch version with a higher CUDA version than your current platform.) everything works as intended.

Now I understand this actually isn't technically the fault of mokuro, but I still think it wouldn't hurt to add a addition to the installation section about ensuring the user gets a PyTorch version that is compatible with the CUDA version they install. If anything, it would save everybody a lot of headaches.

[kenrick95]

When I set this project up for the first time, since it was linked from this project's README, I followed https://xelieu.github.io/jp-lazy-guide/setupMangaOnPC/ and it was quite clear in pointing that out (also on which Python version you should install)