$\boldsymbolλ$-Orthogonality Regularization for Compatible Representation Learning