Measuring What Matters: A Unified Evaluation Framework for GNN Explainability