Combining Group-Contribution Concept and Graph Neural Networks Toward Interpretable Molecular Property Models