Checkpoints assume problem-default dataset

Currently the checkpoint code relies on the user to pass the same `dataset_kwargs["combos"]` used to create the model when resuming training from a checkpoint. If you don't, and are using custom combos, this will result in some hard-to-debug errors.

One workaround is to create a new OPFProblem class and override the `default_combos` property instead of using the `train(dataset_kwargs=...)` API.

The real solution here is to move all the dataset code out of `BasicNeuralNet`, into `BasicNN`.