[avehtari]

In K-fold CV, the cvrefbuilder of get_refmodel.stanreg() and brms::get_refmodel.brmsfit() passes user-supplied latent_ilink, latent_ll_oscale, latent_ppd_oscale and dis to every fold unchanged. The latent vignette writes the negative binomial and Weibull functions as closures over full-data draws (refm_prec, refm_shape) and passes the full-data sigma as dis in the log-normal example. A user who switches these examples to K-fold CV (as in https://discourse.mc-stan.org/t/39275, post 4) therefore scores every fold with full-data draws. Since #568, these draws enter the submodel scores.

This PR documents the behavior and a remedy, without code changes: define the functions inside a function of the fit that calls get_refmodel(), and replace element cvrefbuilder of the reference model object with a function that calls it on the fold refit (get_refmodel.stanreg() and brms::get_refmodel.brmsfit() do not accept a cvrefbuilder argument). It adds a paragraph to ?extend_family, a vignette section "K-fold CV with custom response-scale functions" with the negative binomial example, pointers from the Weibull and log-normal examples, and a NEWS bullet.

I ran the new vignette chunk on a small negative binomial stan_glm() model with 2 folds: with the default cvrefbuilder both folds used the full-data precision draws, with the custom one each fold used its own, and the cross-validated elpd changed by 1 to 5 per submodel size. I did not build the full vignette.

This PR was assisted by Claude.

[aloctavodia]

I added a few minor comments.