Nano: Nested Human-in-the-Loop Reward Learning for Few-shot Language Model Control