PolicySavedModelTrigger has no sweeping ability.

[brianorbrain]

I am having an issue with the saved model trigger when using the learner API. Over a long enough time period, the saved model trigger has no limiter for the number of policy checkpoints it saves. This will eventually completely fill the drive and stop training.

The regular checkpoint saver has the ability to sweep via tf.train.CheckpointManager _sweep function which can prevent an unlimited number of checkpoints from being saved.