Hybrid Explainable Autoencoders for Credit Card Fraud Detection: Integrating Deep Latent Representations with Model-Agnostic Interpretability