Add policy-based model checkpointing callback

[nkrishnaswami]

This PR adds

• PolicyBasedModelCheckpointingCallback to use a policy to perform model checkpointing, optionally including the iteration number to allow saving multiple checkpoints
• ABC CheckpointPolicy to control when to save a checkpoint
• IterationFrequencyCheckpointPolicy to provide the same policy as the existing ModelCheckpointingCallback
• MetricImprovementCheckpointPolicy to save a checkpoint when the target metric improves

Finally, ModelCheckpointingCallback is replaced with a PolicyBasedModelCheckpointingCallback subclass that uses IterationFrequencyCheckpointPolicy.