Comparing information content of representation spaces for disentanglement with VAE ensembles