Currently there's only the _loss_step loss function that's used by L-GMM in lmo.inference.fit. Implementing L-CUGMM would involve creating a _loss_cue. Unlike _loss_step, the _loss_cue loss function updates the weight matrix at each iteration.