Interpretable Classification via Supervised Variational Autoencoders and Differentiable Decision Trees