How Interpretable Are Interpretable Graph Neural Networks?