Easy-to-Hard Generalization: Scalable Alignment Beyond Human Supervision