Mismatched No More: Joint Model-Policy Optimization for Model-Based RL