Inherit information from training data when `train = False` in `GraphDataset` class

For validation and test datasets, the `train` argument is set to `False` and a `dataset_train` must be specified. 

Since attributes such as `target`, `task`, `node_features` and `edge_features` (perhaps others) must be identical between training, validation, and test datasets, it would make sense to read these from `dataset_train` rather than asking the user to specify them again.