Retrospective on `pretraining` functionality

[EvanTrop]

Bugs/feature requests that popped up during the implementation of pretraining! Some should be fixed before merging to main, at least.

• Fix AnnData caching #136
• Pretraining experiment/task config should incorporate fine-tuning tasks/refer to downstream task configs #137
• Make self.partition state unique to each DataLoader worker. #138
• Enable multiple heads/loss functions in a single pretraining run #139