MazeDataset - ensure different representations are always available

[valedan]

Currently it's possible to have a dataset where any of mazes_objs, mazes_tokens, and mazes_array are None. It would be nice if we could always rely on these representations being present.

I think a nice way to do this would be to choose one of these as the "canonical" representation, and have that be the only version that is actually stored. Then the other representations could be provided by functions that convert from the canonical version as needed. This should help tidy up __init__ too as right now there is a lot of code in there for converting between these formats.

[mivanit]

working on this on the zanj-integration-datasets branch (which, misleadingly, contains a rework of the datasets, training loop, and CLI for training).

rough plan:

• integrate with ZANJ for easier saving/loading
• canonical representation of a MazeDataset as a list of SolvedMaze
• preserve the old interface via cached properties
• seamlessly try to get a dataset from local file, download, or generate on the fly (see MazeDataset.from_config() function)
• pass tests, add new ones
• change to new interface in example notebooks and tests

[mivanit]

resolved in #177