Any-step Dynamics Model Improves Future Predictions for Online and Offline Reinforcement Learning