Allow preloaded data

[mdreier]

When running machine learning programs on Google Cloud TPUs (e.g. from Google Colab), the data must be stored on Google Cloud Storage. The current implementation uses methods, especially tf.keras.utils.get_file, which do not support this.

The pull request achieves two things:

• Replace methods that do not support GCS paths by those which do
• Allow callers to specify a path where data is stored. In this case it will not be downloaded

train_data, test_data, classes = ImageWoof.load_data('gs://my-gcs-bucket')

Example

This workbook can be run either on GPU or on TPU.

Possible Improvements

• Replace tf.keras.utils.get_file with another method which allows storage directly into GCS
• Error handling when storage directory does not contain the required data