3-D Magnetotelluric Deep Learning Inversion Guided by Pseudo-Physical Information