Refactor: checkpoint, stopping logic

[RobotSail]

We currently have the library set up to perform actions such as saving checkpoints and finishing training based on how long we've been training for. But the criteria for these can differ in ways such as:

• training for N epochs
• training up to K samples seen
• saving every N epochs OR saving every K samples

There are also different types of checkpoints to save. For full fine-tuning we may export a HuggingFace formatted checkpoint for users to consume, but we may also want to occasionally checkpoint the DeepSpeed state for later resumption of training.

For the reasons listed above, we should refactor the way we communicate the criteria for actions to be simpler and allow for the library to be more predictable.