Bayesian Optimization of a Lightweight and Accurate Neural Network for Aerodynamic Performance Prediction