Port training and evaluation code

[drubinstein]

Beginning the work on bringing over the training and evaluation code. Please point out if I have to remove any keywords that exist.

I have not tested this code. This code will also require a lot of documentation and style adjustments e.g. typing and testing + possibly some generalization beyond GCP. The version of mirdata used may also be slightly out of date and there probably needs to be a better system for downloading the mir datasets.

Theoretically you should be able run the individual pipelines in dataset.py with DirectRunner then run the training script and subsequent evaluation.