Learning disentangled representations for explainable chest X-ray classification using Dirichlet VAEs