Recovering from Out-of-sample States via Inverse Dynamics in Offline Reinforcement Learning