Stabilizing Transformers for Reinforcement Learning