Program Learning with Verifiable Rewards: Symbolic Backpropagation for Post-Training LLMs