`ref_group` should not be limited to displayed split levels

Currently, ref_group appears to require the reference level to be included in the levels retained by split_fun. For example:

table(DM$ARM)

     A: Drug X     B: Placebo C: Combination 
           121            106            129

lyt <- basic_table() |>
  split_cols_by(
    "ARM",
    split_fun = keep_split_levels("A: Drug X"),
    ref_group = "B: Placebo"
  ) |>
  analyze("AGE")

build_table(lyt, DM)

results in:

Error in .add_ref_extras(spl, df, ret) : length(ref_ind) == 1 is not TRUE

From the user perspective, this seems to be an unnecessary and unreasonable limitation. I think ref_group should be allowed to be any valid level of the split variable, even if that level is not displayed or not observed in the data.

This includes:

• A factor level that exists but is not observed in the data.
• A factor level that is not present at all - in this case, perhaps a warning would be appropriate.

Simply put, the split function controls what is displayed, while ref_group should be allowed to specify the reference level independently.

Reproduce the issue using the supplied split_cols_by(), keep_split_levels(), ref_group, and build_table() example. Start at the .add_ref_extras error path and verify that ref_group can refer to an un displayed or unobserved valid level independently of the levels retained by split_fun, with the requested warning behavior for an entirely absent level.