Listening with Attention: Entropy-Guided Explainability for Transformer-Based Audio Models