refactor conll2003

[ArneBinder]

This PR simplifies the usage of the conll2003 dataset and is in preparation for ArneBinder/pytorch-ie-hydra-template-1#1.

Changes:

• add PIEDatasetDictto the data.datasets package, defined as Dict[Union[str, Split], List[Document]] in analogy to huggingface DatasetDict.
• split load_conll2003:
  • load_hf_conll2003: this loads the dataset as an huggingface datasets.DatasetDict
  • from_hf: this converts the output from above to an PIEDatasetDict
• add a helper method single_split_to_dict that unifies the output of huggingface load_data to a (iterable)DatasetDict. This may be moved to another place in the future, if it proves useful.

Motivation for the split: In the future (e.g. the template), we will mostly just need conll2003.from_hf and use methods from huggingface datasets to get the data in the first place.