Automatically copy model training artifacts to cloud storage.

[kramachandran]

Overview

Currently users have to add code to copy training artifacts to cloud storage. This degrades the user experience because requires users to write different code when training locally, as opposed to training remotely via either Kubeflow or GCP hosted services. Further, when executing a full notebook, there is no way to transfer the results to a cloud storage (users cannot add this code to the notebook itself.)

To make good on the Fairing promise of write once/run anywhere the system should be able to automatically move artifacts from user specified directory to a user specified cloud storage location. To further simplify the process, the system can provide a sensible default directory to write artifacts.

[karthikv2k]

CC @martin-gorner