Evaluating the Capabilities of Large Language Models for Multi-label Emotion Understanding