Learning Fair and Interpretable Representations via Linear Orthogonalization