Grasp, See and Place: Efficient Unknown Object Rearrangement with Policy Structure Prior