The Impossibility of Inverse Permutation Learning in Transformer Models