posterior_epred for binomial models different than brms

@bgoodri @paul-buerkner Here's an important different between rstanarm and brms that @lauken13 and I noticed when working on the new MRP package:

Right now rstanarm and brms have different behavior for posterior_epred() for binomial models. In rstanarm we return estimates of p whereas from what I can tell brms seems to return estimates of N * p. Strictly speaking, the brms implementation is more consistent with the name epred in that N*p is the expectation, not p. That said, I find that I pretty much always want p and not N * p when I'm actually using posterior_epred() in practice (e.g., that's what we want for the MRP package). So I think the brms implementation is more faithful to the function name but I find the rstanarm implementation to generally be more useful (at least for MRP and other applications I work on).

How should we reconcile this? It's unfortunate to have different behavior here, but I can see the merits of both.

The issue concerns posterior_epred() for binomial models and contrasts rstanarm with brms, but it names no files or tests. Start by locating the posterior_epred() implementation and reviewing how its binomial return value is defined. Done means reaching and documenting a decision about returning p versus N * p, then aligning the behavior accordingly.