Crafting Interpretable Embeddings by Asking LLMs Questions