Support fine-tuning

Add model_usage: supported to all training workers.

When a model version is provided (self.model_version_id is not null), we should load the model's files, just like we do for the inference workers.

This previous checkpoint should be loaded right before training.