On the Practical Computational Power of Finite Precision RNNs for Language Recognition