A Low-Computational Cost Reduced-Order Neural Network-Based Approach for Accurate Modeling of Two-Phase Flow in Pipes