Support for scikit-learn's LatentDirichletAllocation and NMF, extend support for GaussianMixtureModel and BayesianGaussianMixtureModel

[jeremymanning]

We should add support for LDA and NMF: http://scikit-learn.org/stable/modules/classes.html#module-sklearn.decomposition

These can be supported both via cluster and reduce:

Cluster: behaves similarly to GaussianMixtureModel (estimate how much of each factor is reflected by/in each observation)
Reduce: to reduce the data to n-dimensions, fit LDA/NMF with n components

Some sample code may be found here: https://github.com/ContextLab/storytelling-with-data/blob/master/data-stories/twitter-finance/twitter-finance.ipynb

We could play a similar trick with Gaussian mixture models-- to use a Gaussian mixture model to reduce the data to n dimensions, we could fit a GMM with n components and then use the membership mixing labels as the coordinates.

[jeremymanning]

Status in the 1.0 update (PR #272 #272): the cluster half of this ask is fully done — hypertools/cluster/cluster.py's mixture_models dict already supports GaussianMixture, BayesianGaussianMixture, LatentDirichletAllocation, and NMF, returning per-observation membership-proportion rows that sum to 1 (verified live for all four). The reduce half is not: hyp.reduce(x, reduce='NMF', ndims=3) still raises ValueError because hypertools/reduce/reduce.py's models dict only has classic decomposition/manifold models.

How this could be built on the 1.0 architecture: hypertools/core/model.py's apply_model already solves this — its registry merges reduce.models, cluster.models, and cluster.mixture_models, and _apply_single's mode='auto' logic already picks predict_proba for GMM/BayesianGMM and fit_transform for NMF/LDA. Verified live: apply_model(X, 'NMF', ndims=3) and apply_model(X, 'GaussianMixture', ndims=3, mode='predict_proba') both work today. The concrete fix is to make hyp.reduce() delegate to (or share the registry/dispatch logic with) apply_model, rather than maintaining a separate, narrower models dict — either by adding the four mixture models into reduce.models with the same mixture-handling branch already in cluster.py, or by having reduce() fall back to core.model.apply_model for names it doesn't recognize locally.

Alternatives / prior art:

• No external library gap here — scikit-learn's own NMF/LatentDirichletAllocation/GaussianMixture/BayesianGaussianMixture are exactly what's needed and are already hypertools dependencies; this is purely a wiring gap between hypertools' own reduce() and apply_model().
• The issue's suggestion to "fit a GMM with n components and use membership proportions as coordinates" is exactly the mode='predict_proba' path apply_model already implements.

Tagging as low effort; leaving open for consideration in a future release.