Improving the accuracy and spatial resolution of precipitable water vapor dataset using a neural network-based downscaling method