Train custom classes

[teunbrand]

This PR aims to fix #480.

For discrete training, is.factor(x) is swapped out by !is.null(levels(x)).
That way, any custom class that implements levels.custom() should be able to train a range.

For continuous training, we require the result of range(x) to be numeric (rather than x itself).
Error message is slightly terser, but that isn't too bad.

devtools::load_all("~/packages/scales")
#> ℹ Loading scales
#> Warning: package 'testthat' was built under R version 4.4.2
library(vctrs)

newdisc <- new_list_of(x = list("A", c("B", "C")), ptype = character(), class = "foo")
levels.foo <- function(x) unique(unlist(x))
train_discrete(newdisc)
#> [1] "A" "B" "C"

newcont <- new_list_of(x = list(1, c(5, 3)), ptype = double(), class = "bar")
range.bar <- function(x, ...) range(unlist(x), ...)
train_continuous(newcont)
#> [1] 1 5

^{Created on 2024-11-19 with reprex v2.1.1}

[teunbrand]

The as.numeric() requirement is in L55. Because as as.numeric(LETTERS) returns a vector of NA_real_ without throwing an error, it seemed more sensible to require the output of range() to be numeric,