Prompt caching but for RL – 7.5x speedup on long-prompt/short-response workloads