Remove torch dependency.

[rmlarose]

As far as I know (@naezzell correct me if I'm wrong), we're only using torch to do PCA. Torch is a huge download -- only using PCA from it is a waste. We should avoid this by using a native implementation of PCA, or another way around the torch dependency.

I'm not adding torch to requirements.txt with the expectation that it will be removed.

[naezzell]

It's removed from PCA (reduce_features), but torchvision is still used to get certain data sets. Yousif probably only did it this was since original PCA did use PyTorch already.

[rmlarose]

Is there a better way to get data sets? Which data sets in particular are we getting via torch?