Better API around continuing training

[matt-gardner]

Currently, if you want to load a model trained on one dataset and continue training, you can kind of do that, but it will end up overwriting the model that you loaded, which is a problem. We should add nice functionality around continuing training.