switch Dataset to pandas Dataframe + torch Dataset

[NickleDave]

• lets pandas and keras do the heavy lifting
• makes it possible to work on bigger data sets, avoid loading everything into memory all at once
• don't have to maintain a fancy data structure
• lets tests actually work again -- had non-deterministic failures because of Dask dark magic

[NickleDave]

this ended up becoming a very large "feature branch", #88
Closing because that branch was merged in.
Switched to Pandas DataFrame for datasets, and switched to torch to use their Dataset / Dataloader classes (instead of keras, because of bugs with keras.utils.Sequence)