posterior_survfit.stansurv fails with non-numeric grouping factors

Summary:

posterior_survfit fails for stan_surv models when predicting for new groups (i.e. levels of the random effect) and the original grouping factor in the model was a character or factor data type, and not numeric.

I think the error is occurring here since the object$xlevs include the grouping factors when they were character variables. This isn't ideal because it then prevents new groups (in that case the code interprets them as new "x levels") from being used. I think any grouping factors need to be removed from objects$levs list, or something like that.

Reproducible Steps:

library(rstanarm)

data <- data.frame(
    trt = rep(c("A", "B"), each = 90),
    group_numeric = rep(1:6, each = 30),
    group_string = rep(letters[1:6], each = 30),
    group_factor = factor(rep(letters[1:6], each = 30)),
    status = sample(0:1, 180, replace = TRUE),
    days = sample(10, 180, replace = TRUE))

mod_numeric = stan_surv(
    formula = Surv(days, status) ~ trt + (1 | group_numeric),
    data = data,
    basehaz = "weibull",
    chains = 1,
    iter = 100,
    cores = 1)

mod_string = update(
    mod_numeric, 
    formula = Surv(days, status) ~ trt + (1 | group_string))

mod_factor = update(
    mod_numeric, 
    formula = Surv(days, status) ~ trt + (1 | group_factor))

nd <- data.frame(
    trt = c("A", "B"), 
    group_numeric = 999, # new value in the prediction data
    group_string = "zzz", # new value in the prediction data
    group_factor = factor("zzz")) # new value in the prediction data

# works, with warnings about not being factors
ps_numeric = posterior_survfit(
    mod_numeric, 
    newdata = nd,
    type = "surv",
    times = 0)

# doesn't work, doesn't allow new group levels
ps_string = posterior_survfit(
    mod_string,
    newdata = nd,
    type = "surv",
    times = 0)

# doesn't work, doesn't allow new group levels
ps_factor = posterior_survfit(
    mod_factor,
    newdata = nd,
    type = "surv",
    times = 0)
```

Start with posterior_survfit and the grouping-level handling in R/pp_data.R around line 374. Run the reproducible example using character and factor grouping variables, comparing it with the numeric case. Trace why grouping levels are treated as x levels, and consider the issue's suggested separation; done means predictions for new non-numeric groups work without the current failure.