Several issues about TUDataset

I encountered several usability issues when developing the benchmark code on TUDataset.

1. There are two datasets `LegacyTUDataset` and `TUDataset`. They are quite different -- `TUDataset` does not have node features while `LegacyTUDataset` does. We should merge them and provide options to include node features per user request.
2. Does not support slicing or list indexing, which is very useful during train/val/test split. Ideally, `dataset[index_list]` should return a sub-dataset. I imagine this might be a common issue for all graph classification datasets.
3. Does not support shuffling. We should have sth like `dataset.shuffle()` which returns a new shuffled dataset, which is useful for creating stratified train/val/test splits.
4. `dataset.num_labels` returns a `numpy.ndarray` instead of an integer.
5. Lack an API for getting the number of input features, e.g., `dataset.num_features`.