Turing Completeness of Bounded-Precision Recurrent Neural Networks