The Art of Scaling Reinforcement Learning Compute for LLMs