Upload is controlled by a new option `logger`/`wandb`/`log_model`. When
enabled, model files are uploaded to W&B as artifacts attached to the
current run, at the frequency specified by `save_checkpoint_freq`.

I'm using this feature for several purposes:
- Centralized storage of models being trained in various locations.
- Decoupling of training and testing. I usually upload models as training
progresses, while a job running in a different computer downloads them
and performs arbitrary metrics / evaluation tasks.