`..new` is calculated wrong in lencode steps

The unseen levels are calculated based on the mean of the coeficients rather than the mean of global data. This should be fixed to better reflect the literature.

Make sure that the documentation is changed accordingly.

This change will be easily backward compatible as it changes how new values will change only.

data <- data.frame(
  outcome = rnorm(1000) + c(rep(10, 900), rep(0, 100)),
  predictor = c(rep("Big", 900), rep(letters[1:10], each = 10))
)

library(tidyverse)

data |>
  count(predictor)
#>    predictor   n
#> 1        Big 900
#> 2          a  10
#> 3          b  10
#> 4          c  10
#> 5          d  10
#> 6          e  10
#> 7          f  10
#> 8          g  10
#> 9          h  10
#> 10         i  10
#> 11         j  10

data |>
  summarize(
    mean = mean(outcome),
    .by = predictor
  )
#>    predictor        mean
#> 1        Big  9.92621834
#> 2          a -0.12884918
#> 3          b  0.24802560
#> 4          c  0.12339453
#> 5          d  0.33307724
#> 6          e  0.08705590
#> 7          f  0.86433875
#> 8          g  0.42452332
#> 9          h  0.42548890
#> 10         i -0.07257279
#> 11         j -0.67403943

embed:::glm_coefs(y = select(data, outcome), x = pull(data, predictor))
#> # A tibble: 12 × 2
#>    ..level ..value
#>    <chr>     <dbl>
#>  1 a       -0.129
#>  2 b        0.248
#>  3 Big      9.93
#>  4 c        0.123
#>  5 d        0.333
#>  6 e        0.0871
#>  7 f        0.864
#>  8 g        0.425
#>  9 h        0.425
#> 10 i       -0.0726
#> 11 j       -0.674
#> 12 ..new    0.256

mean(data$outcome, trim = 0.1)
#> [1] 9.717217

Start at the embed:::glm_coefs entry point shown in the example and trace how the ..new level is calculated. Use the supplied unbalanced-data example to verify that ..new reflects the global trimmed mean rather than the mean of coefficients, then update the related documentation accordingly.