Train Once and Explain Everywhere: Pre-training Interpretable Graph Neural Networks