Implement Batch size restrictions

[brechtmann]

To solve the Basenji Issue in the long run we could think of implementing a batch size restriction.

This could be an option in the model.yaml file, which is then handed over to the respective data-loader. It could also be specified in the default data loader section:

default_dataloader: 
  defined_as: kipoiseq.dataloaders.SeqIntervalDl
  default_args:
    auto_resize_len: 131072
    ignore_targets: True  # don't return any output labels using the bed file

[Avsecz]

I wouldn't patch kipoiseq.dataloaders.SeqIntervalDl as it would look weird in the kipoiseq codebase. Instead, we should just add an extra flag to the model.yaml to pose a batch size restriction and then print an informative error message when the user tries to use a different batch size. It's really a model issue not a dataloader issue.