Towards Mechanistic Interpretability of Graph Transformers via Attention Graphs