Add train and eval mode to taskmodules

[ChristophAlt]

Problem: Some taskmodules require different behavior during training.
Solution: Introduce a training flag (controlled by .train(..) and .eval(..)) similar to pytorch's Module, as users are already familiar with the usage.

[ArneBinder]

After thinking about this for a while, I'm not so much in favor of it since this introduces a flag that can be set in arbitrary locations and its state may not easy to track. I would propose another solution: a method Taskmodule.prepare_documents that is called before Taskmodule.encode_input and by default does nothing. However, it gets a boolean parameter ground_truth_available which is set to the value of encode_target. This method can be used to e.g. add missing annotation layers or some metadata to the document. It can also be used to e.g. add candidate entity pairs for RE with bit more sophisticated logic (this would facilitate some parts of the SAM models). See #102 for the required changes.

[ArneBinder]

This requires the bugfix PR #108 to be functional.