Add a .locale argument to arrange() and use radix ordering

[DavisVaughan]

Closes #4962
Closes #5090
Part of #5808

Requires this vctrs PR r-lib/vctrs#1375

The above vctrs PR switches vec_order() to always use radix ordering. This propagates through to dplyr in 3 places:

• arrange()
• group_by()
• with_order()

This PR tackles arrange() by providing a new .locale argument with 3 possibilities:

• NULL, the default, for C locale
• A single string, like "en_US", to use a specific locale, backed by and requiring stringi
• A formula/function to generate the sort key (i.e. stringi::stri_sort_key(), or tolower())

This PR is currently focused on the implementation, I've left TODOs for the news bullet and for documentation of locale handling.

The snapshot tests are a little noisy. Dev vctrs requires dev rlang, which tweaked how deparsing with ~ works, changing a few snapshot tests here.

Build errors seem to be from the fact that rlang now exports ellipsis functions, causing some warnings

---

This would be a meaningful breaking change, as the system locale is no longer being respected. However, there are multiple benefits from this change:

• Sorting character vectors in the C locale is much much faster
• Even generating a sort key and then sorting in the C locale is faster than the current approach
• This removes dependence on an environment variable (LC_COLLATE), which we generally advocate in favor of explicitly specifying arguments. In other words, this gets us better consistency across platforms.
• This increases consistency with data.table, who always uses the C locale
• This increases consistency with stringi (and you could make an argument for stringr too)

For English, the biggest change is that uppercase letters now sort before any lowercase letters. Previously it used a natural ordering of c("a", "A", "b", "B").

For other languages, sorting directly in the C locale is often not very meaningful, so they would have to supply a locale identifier.

[hadley]

LGTM — we'll just need to leave this sit until we figure out a unified strategy.

[DavisVaughan]

Updated so that:

• .locale = NULL defaults to "en" if stringi is available, otherwise it falls back to the C locale with a warning
• .locale = "C" is now a valid option

[DavisVaughan]

Updated so that:

• We have a new exported dplyr_locale() helper. It defaults to "en" if stringi is installed, otherwise it falls back to "C" with a warning. Users can set options(tidyverse.locale_collation) if they want to globally override the default locale.
• .locale in arrange() now defaults to dplyr_locale()
• .locale no longer allows NULL
• The scoped variants of arrange() have gained .locale to ease the transition a little

[DavisVaughan]

Closing in favor of the cleaner #5942