Allow 0-column newdata if positive number of rows

[jgabry]

fixes #491

This PR avoids performing one of the checks on newdata that resulted in an error if newdata had 0 columns. Allowing 0-column newdata is useful for doing posterior_predict() for intercept only models, as pointed out in #491 by @davidkane9. With this change, the following now works:

fit <- stan_glm(mpg ~ 1, data = mtcars)
posterior_predict(fit, newdata = tibble::tibble(.rows = 1)) # generate predictions for 1 obs
posterior_predict(fit, newdata = tibble::tibble(.rows = 2)) # generate predictions for 2 obs
...

[jgabry]

This also works with posterior_epred() and posterior_linpred(). I will also add a test.