Provide better interface or documentation for per-geo modeling

[brookslogan]

Context

@dsweber2 was just noting that

epi_df %>%
  epi_slide(
    ~ .x %>%
      group_by(geo_value) %>%
      arx_forecaster(.....)
    )

doesn't fit per-geo models; it actually just ignores the grouping altogether.

We suggested "transposing" the operations, but @rnayebi21 found that

epi_df %>%
  group_by(geo_value) %>%
  epi_slide(~ arx_forecaster(.x, .....), .....)

doesn't work either; .x doesn't have the geo_value column thus lacks epi_dfness. I believe these problems also apply to when you are trying to do version-faithful backtesting with epix_slide().

Workarounds seem a little bit of a pain, either

• fixing up the first approach by doing something like
  • split + map + bind_rows, or
  • group_by + group_split + map + bind_rows, or
  • group_by + group_modify(.keep = TRUE), or
  • group_by + reframe (using the deprecated-but-not-replaced cur_data_all()...)
• fixing up the second approach by reconstructing an epi_df inside the slide computation using .x, .group_key, and .ref_time_value, or
• [mutate(geo_value2 = geo_value) and group by that instead of geo_value. Or just group_by(geo_value2 = geo_value).]

The first workaround seems more modular (you can have a list of forecasters that can all rely on ungrouped slides, rather than having to do a different type of slide call for each one).

Proposal

1. Make arx_forecaster() check specifically if there's a missing geo_value and hint that if they were doing a grouped epix_slide() or epi_slide() with geo_value in the group variables, that won't work, and to do <some workaround / feature> instead.
2. Check if input to arx_forecaster() etc. is grouped; if so, either
   • warn
   • abort
   • fit & forecast one model per group
3. [Also check for groupedness and warn/abort in the epi workflow internals.]

Musings

We can also probably make things easier epiprocess-side, by adding a .keep parameter if we're not already able to forward to group_modify() via dots. But I'm not sure we actually want to... this makes it easier to use epi_slide() for forecasting when it shouldn't actually be (epix_slide() should be favored and maybe renamed to make this clear).

[@dshemetov points out we should document this geo-grouped epi_slide gotcha in epiprocess. And actually fixing what's going wrong is part of a much larger project, epiprocess#223.]

[dajmcdon]

• I think this is possibly outside the scope of arx_forecaster(). Just make a new one.

Maybe this was described above, but I'm not quite clear. It sounds like the major issue is that calling group_by() on an epi_df has a strange effect on the geo_value (and whatever you are grouping on) that causes it to behave poorly with the epi_workflow processing. Is that right?