limiting loader should be able to only limit the training data

Currently, the limiting loader is able to restrict how many documents per target are being returned by the dataloader. It could be useful to only restrict the training data, as more testing data will only increase the certainty of the evaluation, not the outcome.