ProtGNN: Towards Self-Explaining Graph Neural Networks