Mechanistic Interpretability of Emotion Inference in Large Language Models