On the Expressiveness of Rational ReLU Neural Networks With Bounded Depth