ENH: helper: ordered multinomial logit parameter transformation for discrete mixture

A simplified version of multinomial logit with monotone choice parameterization.

e.g. we have 5 classes (discrete points on the real line) and we don't know the location of the points nor the probabilities. We want to estimate a discrete mixture distribution for a sample or as one component (mixing probability) in a larger model.

To get a unique parameterization of the "states", we can impose monotonicity x0 < x1 < ... < xk where xi are the support points.

parameterization c0, c1, ..., ck such that
x0 = c0
x1 = x0 + np.exp(c1)
...
xk = x{k-1} + np.exp(ck)

This is what I used for ordered Logit/Probit, but is more widely useful, for example to implement mixed effects models with finite support mixing distribution.

Start by locating the existing ordered Logit/Probit parameterization referenced in the issue and study how its monotone support transformation is represented. Define the helper’s scope for ordered support points and mixture use cases, then add tests covering the five-class example, monotonicity, and parameter recovery; done means the transformation is usable for discrete mixture models.