MELT: Towards Automated Multimodal Emotion Data Annotation by Leveraging LLM Embedded Knowledge