Improving Molecular Modeling with Geometric GNNs: an Empirical Study