Automatically unpack in `across()`

[hadley]

To support this sort of functionality:

library(dplyr, warn.conflicts = FALSE)

quantile_df <- function(x, probs = c(0.25, 0.5, 0.75)) {
  tibble(
    "{{x}}_val" := quantile(x, probs),
    "{{x}}_quant" := probs
  )
}

df <- tibble(
  grp = rep(1:3, each = 10),
  x = runif(30),
  y = rnorm(30)
)

df |> 
  group_by(grp) |> 
  summarise(across(x:y, quantile_df))

[DavisVaughan]

See also #5563

[DavisVaughan]

UI proposed by @lionel- and I

Essentially the idea is to add one more layer on top of what we already get back from across(cols, fns, .names = spec)

@param .unpack If `fns` returns a data frame, should that data frame be unpacked?
  - If `TRUE`, uses a `"_"` separator to separate the names between the original column/function name combination and the data frame column names. This is identical to using `"{outer}_{inner}"` (see below).
  - If `FALSE`, no unpacking is done.
  - A glue specification can also be used, using the keywords `"{outer}"` and `"{inner}"`. `outer` represents the names formed by the `.names` argument, and `inner` represents the names of the data frame returned by `.fns`

# Default
# Same as:
# `.names = "{col}_{fn}", .unpack = "{outer}_{inner}"`
across(x:y, quantiles, .names = NULL, .unpack = TRUE) 

# Don't unpack 
across(x:y, quantiles, .unpack = FALSE)

# Unpack with a specific glue spec
across(x:y, quantiles, .unpack = "{outer}.{inner}")

[hadley]

I assume you discussed and rejected adding another variable to the existing name specification? i.e. across(x:y, quantiles, .names = "{outer}.{inner}.{fn}")? Or do we assume that when unpacking there's only ever a single .fns? (It's hard to imagine having two data frame functions returning the same number of rows, but I guess it might happen?)

[DavisVaughan]

Note: We also realized that we'd probably have to default it to unpack = FALSE. I think it is impossible to do across() expansion with unpack = TRUE, so we'd only ever attempt it when it is FALSE

[DavisVaughan]

I think having the additional argument is nice because we are going to need it regardless to be able to opt-in to the unpacking (or out if we figure out a way to allow it to default to unpack = TRUE).

I think putting the glue spec in .unpack is nice because it allows people to mix functions that need unpacking with those that don't:

across(
  x:y, 
  list(lag = ~multilag(.x, 1:5), double = ~.x * 2), 
  .names = "{col}.{fn}",
  .unpack = "{outer}.{inner}"
)

In this example when x.lag and y.lag are unpacked further, they use the .unpack glue spec, but x.double and y.double ignore it.

It's hard to imagine having two data frame functions returning the same number of rows, but I guess it might happen?

multilag() returns a data frame with the same number of rows as the input, and I could probably see other functions that do that too. i.e. someone might do multilag() combined with a multilead() in the same across() call.