A Framework for Variational Inference of Lightweight Bayesian Neural Networks with Heteroscedastic Uncertainties