Discriminative Feature Attributions: Bridging Post Hoc Explainability and Inherent Interpretability