Learning Theory Can (Sometimes) Explain Generalisation in Graph Neural Networks