Mercury-ml

- Custom __str__ functions have been added to DataSet and DataBuch so that printing instances of these will show a sensible hierarchy of information
- A generic "read_data_bunch" function has been added to mercury_ml.common.tasks which can take arbitrarily named DataSets
- The `read_pandas_data_set` function in `mercury_ml.common.source_reading.disk` has been made more abstract. It can will now create arbitrarily named data_wrappers based on what is input in the `data_wrappers_params` field

- Saving to S3 now always overwrites as the methods used so far to check whether an object exists has proven to be less than reliable

- Fixed a small error where checking whether a S3 object already exists did not work correctly

- This release sees the introduction to TensorFlow 2.0, for now using the pre-release alpha version. The `keras` implementations have all been replaced with `tensorflow.keras`.
- It is now also possible to save models using the tensorflow saved_model API
- The "providers" / "containers" structure has been flattened. Modules that were previously under `mercury_ml.common.providers` and under `mercury_ml.common.containers` are now simply under `mercury_ml.common`. The same goes for the structure under `mercury_ml.<machine learning engine>`

- Under `mercury_ml.keras.providers.model_fitting` bot `fit` and `fit_generator` now:
1. Accept `**kwargs` to be passed to the underlying Keras functions
2. If `save_best_model=True` will only reload model weights (instead of importing an entire model as was done before). This keeps model training history intact

- MongoDBSingleton now only imports pymongo from within the __init__ scope
- saving artifacts to MongoDB will now correctly update only the lowest level the the document_key provided)
- copying from s3 to disk will now automatically create the local directory to be downloaded to