use cuda if available, speeding up training

[gantrol]

Total training time on CPU(13700F): 138.42 seconds
Total training time on GPU(3080Ti): 64.58 seconds
GPU speedup: 2.14x

[karpathy]

So the tricky thing here is that using GPU (or MPS) could actually be slower because our networks here might be very very small. So the auto-detect logic is a bit more complex possibly. Hmm let me think.

[karpathy]

This is what I have
#14

Not 100% sure how to handle this right now. It's introducing a layer of complexity that I'm not sure I want to "bundle" into this module.