On the Dimension of Pullback Attractors in Recurrent Neural Networks