Return per-category and per-threshold estimates for ordinal models with the emmeans backend

[jmgirard]

This PR goes into support_emmeans_ordinal (#669), not into main.

With #669, estimate_means(m, "Type", predict = "prob", backend = "emmeans") returns 0.333 in every row. For ordinal modes, emmeans adds a pseudo-factor to the reference grid. If that factor is not in specs, emmeans averages over it. This PR adds that factor to specs:

• For predict = "prob", the response factor is added. The output has one row per response category, in a Response column.
• For "cum.prob", "exc.prob" and "linear.predictor", the cut factor is added. The output has one row per threshold, in a Threshold column.
• The estimate columns are Probability, Mean_class, Latent and Linear_predictor. These names are also added to the list of known estimate names, so coef_name is set.
• emmeans names the response factor after the left-hand side of the formula, for example factor(y). The PR uses the same name.

For ordinal models (not brms), the default predict on the emmeans means path is now "prob". Before, the default was "response", which emmeans ignores for these models, so latent values were printed as "Probability". estimate_contrasts() and estimate_slopes() are not changed. A test pins their current output for polr.

Tests in test-ordinal.R:

• predict = "prob" and the default, for one and two focal terms, compared with the marginaleffects backend (polr, clm), within 1e-6.
• mean.class and latent, and the three threshold modes, compared with emmeans::emmeans() directly (polr).
• A response written as factor(SatNum) in the formula.

The glmmTMB block keeps the version skip (needs glmmTMB >= 1.1.15.2). The NEWS entry and the predict docs are updated.

Found but not changed here:

• estimate_means(m, "Type", backend = "emmeans", mode = "latent"), with mode passed through ..., fails with a memory error. It fails in the same way on main.
• For nnet::multinom with predict = "prob" and the emmeans backend, every row is still 0.33, because only ordinal models get the pseudo-factor.
• test-brms.R:26 fails at random, also on main.