On the Non-Existence of a Universal Learning Algorithm for Recurrent Neural Networks