Can't use init = 0 when >2 group specific terms in stan_{g}lmer

Summary:

stan_{g}lmer always returns an error when init = 0 and there are > 2 group-specific parameters.

Description:

This is only a small bug and you may already be aware of it. Using init = "random" solves the problem. The bug seems to be caused by the Stan make_theta_L function. When init = 0 and there are > 2 group-specific parameters, then make_theta_L returns a vector which contains some NaN values at initialisation.

Reproducible Steps:

library(lme4)
library(rstanarm)
m1 <- stan_lmer(Reaction ~ poly(Days, 2) + (poly(Days, 2) | Subject), sleepstudy, init = 0)

And I seemed to get a similar error on other example datasets.

RStanARM Version:

2.15.3, downloaded from CRAN

R Version:

R 3.4.1

Operating System:

Windows 10

Run the supplied stan_lmer reproducible example with init = 0 and compare it with init = "random". Start by inspecting the Stan make_theta_L function and its initialization inputs; done means the example no longer produces NaN values or the initialization error with more than two group-specific parameters.