Seeing is Believing: Emotion-Aware Audio-Visual Language Modeling for Expressive Speech Generation