Provably Efficient Offline Reinforcement Learning in Regular Decision Processes