Notebook: add pyHealth dataset parsing (incomplete)

Add a custom wrapper around the MIMIC3Dataset pyHealth class that
add functions to process the D_* text tables to extract the natural
language we need for DescEmb.

Processing functions are not complete. TODO actually write these.

Force MIMIC3Dataset to load from local files using global `DATA_DIR_`

Added skeleton for 2 pyHealth tasks readmission_pred_task, diagnosis_pred_task.
These are from the tutorial and they will be necessary to take the output of
MIMIC3DatasetWrapper and turn it into a test/train dataset.
TODO - incorporate the text data into the output of these tasks

TODO - hook up BERT model to some fake pasted strings of what we expect the
dataset to look like to verify it runs as we expect.