Add a .method argument to the data frame method of arrange()

[DavisVaughan]

Closes #4962

This PR adds a .method argument to the data frame method of arrange(). It defaults to "dplyr_auto", but can also be any valid order() method.

"dplyr_auto" is more aggressive than "auto" in two ways:

• In addition to integer, numeric and logical vectors, it also uses radix ordering for character vectors. This forces the use of a C-locale, but is very fast. This is a breaking change for the defaults of arrange().

• It uses typeof() rather than is.*() functions when determining if radix ordering can be used to operate at a level lower than S3 dispatch. This probably doesn't matter too much, because it is called on the comparison proxies of the columns (which generally returns the underlying raw data), but it still felt more correct.

library(dplyr, warn.conflicts = FALSE)

set.seed(123)

size <- 1000000

df <- tibble(
  a = sample(c("a", "b", "c", "d"), size, replace = TRUE),
  b = sample(1:20, size, replace = TRUE)
)

# Master
bench::mark(arrange(df, a, b), iterations = 20)
#> # A tibble: 1 x 6
#>   expression             min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>        <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 arrange(df, a, b)    1.99s    2.22s     0.449    41.1MB     1.80

# This PR
bench::mark(arrange(df, a, b), iterations = 20)
#> # A tibble: 1 x 6
#>   expression             min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>        <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 arrange(df, a, b)   34.3ms   37.8ms      24.8    48.8MB     105.

[DavisVaughan]

Problems:

1. R 3.2 did not have a method argument at all, so it was forcibly chosen for you. Radix sorting still seems to have been done, but in a very small set of cases ("Method "radix" is only implemented for integer x with a range of less than 100,000")

2. R 3.3 does not have a method = "auto" option, and seems to default to "shell" unless explicitly specified as "radix".

[romainfrancois]

LGTM, but that can equally wait for 1.1.0 ?

[romainfrancois]

Is the first problem enough justification to start depending on R > 3.3 ?

[DavisVaughan]

My current impression is that we will wait on this and maybe eventually add a custom ordering function to vctrs that works on R 3.2 and uses radix sorting, along with the ability to specify ascending/descending per column, but I might be wrong. It seems like that will be a tough project though

[DavisVaughan]

Superseded by #6263