Understanding Privacy Risks of Embeddings Induced by Large Language Models