Spatial Action Unit Cues for Interpretable Deep Facial Expression Recognition