Liquid Time-constant Recurrent Neural Networks as Universal Approximators