Scalable AI Safety via Doubly-Efficient Debate