Globally Convergent Offline Reinforcement Learning with Smoothed Bellman Residual Minimization