Single-draw bug fix for augmented-data and latent projection

[fweber144]

This fixes a bug for augmented-data and latent projection when using a single projected draw for performance evaluation in cv_varsel() with cv_method = "LOO" and validate_search = FALSE. Reprex:

data(inhaler, package = "brms")
inhaler$rating <- as.factor(paste0("rtg", inhaler$rating))
inhaler$subject <- scale(inhaler$subject)[,]
rfit <- rstanarm::stan_polr(
  rating ~ period + carry + treat + offset(subject),
  data = inhaler,
  prior = rstanarm::R2(location = 0.5, what = "median"),
  chains = 1,
  iter = 500,
  seed = 1140350788,
  refresh = 0
)

devtools::load_all()

# Augmented-data projection:
refm_aug <- get_refmodel(rfit)
cvvs_aug_no_refit <- cv_varsel(refm_aug,
                               validate_search = FALSE,
                               nclusters = 1,
                               refit_prj = FALSE,
                               nterms_max = 2,
                               seed = 46782345)
cvvs_aug <- cv_varsel(refm_aug,
                      validate_search = FALSE,
                      nclusters = 1,
                      nclusters_pred = 1,
                      nterms_max = 2,
                      seed = 46782345)

# Latent projection:
refm_lat <- get_refmodel(rfit, latent = TRUE)
cvvs_lat_no_refit <- cv_varsel(refm_lat,
                               validate_search = FALSE,
                               method = "L1",
                               refit_prj = FALSE,
                               nterms_max = 2,
                               seed = 46782345)
cvvs_lat <- cv_varsel(refm_lat,
                      validate_search = FALSE,
                      method = "L1",
                      nclusters_pred = 1,
                      nterms_max = 2,
                      seed = 46782345)

This bug is fixed here by adding drop = FALSE to the relevant matrix subsetting operations. In theory, a similar bug could occur for analogous reference model quantities (that's why drop = FALSE is added there, too), but this is only a theoretical problem as a reference model with a single draw is unlikely to occur in practice.

A NEWS.md entry is still missing here. I will add it as soon as #511 (or #508) has been merged. I would propose the following text for such a NEWS.md entry (under "Bug fixes"):

* Fixed a bug that caused an error when using the augmented-data or latent projection in combination with a single projected draw for performance evaluation in `cv_varsel()` with `cv_method = "LOO"` and `validate_search = FALSE`. (GitHub: #512)