Caching makes saving and loading more difficult

[EvanKomp]

In fact, any functionality that relies on the metadata folder breaks if using joblib.save and load on the model. It does not save the metadata folder (eg. with cache results or subprocess params)

We should overwrite sklearn save and load methods to zip up metadata folder and fumped model, and when loader the model, do some path handling to assign the metadata folder param within the model to the appropriate path.