Add PyTorch loader

The code exists, it just needs to be cleaned.

There might be some work to re-structure drivers in order to do as much ops in Spark (`toPandas` is called prematurely right now, because of inhomogenous python version on the testing cluster when it was implemented.