Dataset fixtures

[wiederm]

We need to clean up the fixtures defined in conftest.py; I propose the following structure:

We have a central place (maybe modelforge.dataset and modelforge.potential) in which we implement a dictionary for the implemented nnps and datasets, in which the name of the datset maps to the class of the dataset (this is already implemented).

We have fixtures with the following base structure:

@pytest.fixture
def dataset_factory():
    def create_dataset(dataset_name:str, for_unit_testing: bool = True, batch_size=64):
        return initialize_dataset(
            dataset_name, for_unit_testing=for_unit_testing, batch_size=batch_size
        )

    return create_dataset
    
    
def initialize_dataset(
    dataset_name:str,
    for_unit_testing: bool = True,
    batch_size: int = 64,
    splitting_strategy:SplittingStrategy = FirstComeFirstServeSplittingStrategy() 
) -> DataModule:
    """
    Initialize a dataset for a given mode.

    data_module = DataModule(
        dataset_name,
        splitting_strategy=splitting_strategy,
        split_file=split_file,
        batch_size=batch_size,
    )
    data_module.prepare_data()
    data_module.setup()
    return data_module

and for each test that should iterate over the datasets and, e.g., different batch sizes, we explicitly add @pytest.mark.parametrize

[wiederm]

This should be done after we merge PR #112