Robust Offline Imitation Learning Through State-level Trajectory Stitching