Securing GNNs: Explanation-Based Identification of Backdoored Training Graphs