Trajectory Graph Learning: Aligning with Long Trajectories in Reinforcement Learning Without Reward Design