Consider adding `grouped_okay` parameter to `is_epi_archive`

[brookslogan]

From #644 (comment)

Previously, we had is_epi_archive with a something like a grouped_okay arg deciding whether to count grouped archives or not. Does that sound useful?

[dshemetov]

Alternatively, can we make grouped archives have both epi_archive and grouped_epi_archive like tibbles

r$> tibble(a = 1, b = 1) %>% group_by(a) %>% class()
[1] "grouped_df" "tbl_df"     "tbl"        "data.frame"

[brookslogan]

We/I originally decided against that because it makes grouped archives inherit ungrouped archive implementations, which can be error prone. Unless we & any other archive method implementers make grouped impls / groupedness checks for everything that at least abort, then a user may get an ungrouped op [or even worse, data structure corruption] expecting a grouped one. I ran into this at some point with grouped_df + add_{row,column} (though in its current state at least, there are checks / signs that might prevent this).

idea: something sort of similar that might be safer is having

• c("grouped_epi_archive", "epi_archive")
• c("ungrouped_epi_archive", "epi_archive")

Haven't really thought this through though. (Potentially related: it would be nice to have an interface / abstract class sort of thing for epi_archives that would allow multiple data structure implementations; e.g., the current diffing thing vs. the disk snapshot thing vs. duckdb vs. a lazy merge etc.)

[It also seems more that ungrouped archives could inherit in some sense from grouped ones, since they are like having a single group. But that might look confusing class-wise.]

[dshemetov]

Could also put is_epi_archive and is_grouped_epi_archive on the same ref page, so the user can OR the two together if they want to allow both.