Generative augmentation and early time-series prediction of gas concentrations: a Beta-Conditional Variational Autoencoder–Gated Recurrent Unit approach