projpred:::auc() incorrect for binomial family with > 1 trials

[fweber144]

Following upon #328, the following shows that projpred:::auc() gives an incorrect result in case of a binomial family with > 1 trials:

set.seed(6834)
nobs <- 100L
mu <- binomial()$linkinv(-0.42 + rnorm(nobs))
trials <- sample.int(30L, size = nobs, replace = TRUE)
y <- rbinom(nobs, size = trials, prob = mu)
dat <- data.frame(y = y, mu = mu, trials = trials)

( auc_val_projpred <- projpred:::auc(cbind(y, mu, trials)) )
## Gives: 0.4909168
# Convert to augmented (long) format:
dat_long <- do.call(rbind, lapply(seq_len(nobs), function(i_short) {
  data.frame(y = c(rep(0L, trials[i_short] - y[i_short]),
                   rep(1L, y[i_short])),
             mu = mu[i_short])
}))
( auc_val_projpred_long <- projpred:::auc(cbind(dat_long$y, dat_long$mu, 1)) )
## Gives: 0.7301717
library(pROC)
auc_val <- auc(y ~ mu, data = dat_long, direction = "<", algorithm = 1)
( auc_val_pROC <- auc_val[seq_along(auc_val)] ) # Indexing only to drop attributes.
## Gives: 0.7301717
stopifnot(isTRUE(all.equal(auc_val_pROC, auc_val_projpred_long, tolerance = 1e-15)))

The value of 0.4909168 for auc_val_projpred should also be 0.7301717. A possible fix would be to convert the data to long format inside of projpred:::auc(), but perhaps there is a less expensive solution in terms of computational efficiency.