Consider using a new distributed batch training logic

[Innixma]

When fitting all the models for a repository, we originally used AutoMLBenchmark (AMLB). This is because AMLB has a well-tested logic for running AutoGluon. However, AMLB is somewhat intimidating to use for those unfamiliar, and thus to increase adoption of TabRepo in terms of researchers fitting their models on our datasets to extend our Repos, it might be better to have a more generic package for fitting models. Particularly one that is simpler to debug and test locally (which AMLB is unfortunately lacking in).

I have seen multiple papers running their methods on TabZilla, so it might be worth understanding what the extension process is in TabZilla and if we can apply a similar process.