Extendable and invertible manifold learning with geometry regularized autoencoders