Reinforcement learning under temporal logic constraints as a sequence modeling problem