Grokking Deep Reinforcement Learning