Flow to Control: Offline Reinforcement Learning with Lossless Primitive Discovery