Disentanglement metrics don't seem to work with continuous factors

[asross]

I might be misunderstanding (e.g. perhaps all the ground-truth factors are meant to be discretized), but it seems like all of the metrics require latent factors to be discrete, e.g. using a classifier rather than a regressor for DCI. This is a bit inconvenient for users who would like to use your library to evaluate disentanglement solely given matrices of continuous latents and ground truth factors (without using the representation_function API). I know it's a design choice rather than a bug, but I'd still like to make the suggestion that it would be helpful for future users if you supported this use case!