Scaling laws for single-agent reinforcement learning