handle split parameter for DatasetDict.load_dataset()

[ArneBinder]

With this PR, it is possible to call DatasetDict.load_dataset(..., split=SPLIT) or DatasetDict.from_json(..., split=SPLIT) to just load the data of a selected SPLIT. This is helpful when using large datasets where getting the data and calling _generate_document() already takes a lot of time and should not be executed for splits that are not used at all.

Usage:

• re-install pytorch-ie: pip uninstall pytorch-ie -y && pip install git+https://github.com/ChristophAlt/pytorch-ie.git@load_dataset_with_split
• call the eval command with +dataset.input.split=test

Changes:

• DatasetDict.load_dataset() and DatasetDict.from_json(): unwrap split parameter and use it if base load_dataset() returns a simple (Huggingface) (Iterable)Dataset instead of a ...Dict
• add type checks to DatasetDict.load_dataset()
• DatasetDict.from_hf(): simplify to only work with (Iterable)DatasetDicts or dicts containing (Iterable)Datasets