Learning Interpretable Style Embeddings via Prompting LLMs