Rewrite ngram implementation with an eye for cpp11 performance

[DavisVaughan]

I ended up doing a full rewrite so you can see what I did to directly fill the output vector on each ngram iteration rather than creating and returning a temporary vector each time. And so you can see the style i use for C++

It is probably "good enough" for now, but will get better when r-lib/cpp11#299 is fixed

Knowing when to unwind_protect() and when to avoid expensive (but safe) cpp11 wrappers does make this pretty challenging to get right.

With CRAN cpp11 (0.4.3)

library(friends)

friends_data <- friends::friends |>
  dplyr::slice(1:10000) |>
  dplyr::pull(text) |>
  stringr::str_split(" ")

bench::mark(
  new = cpp11ngram:::cpp11_ngram(friends_data, n = 2L, n_min = 1, delim = "_"),
  old = textrecipes:::rcpp_ngram(friends_data, n = 2L, n_min = 1, delim = "_"),
  iterations = 50
)
#> # A tibble: 2 × 6
#>   expression      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 new          53.2ms   59.6ms      16.5    1.43MB    0.337
#> 2 old          30.5ms   34.2ms      29.2    1.84MB    1.87

With r-lib/cpp11#299, where the nested unwind_protect() bug is fixed:

bench::mark(
  new = cpp11ngram:::cpp11_ngram(friends_data, n = 2L, n_min = 1, delim = "_"),
  old = textrecipes:::rcpp_ngram(friends_data, n = 2L, n_min = 1, delim = "_"),
  iterations = 50
)
#> # A tibble: 2 × 6
#>   expression      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 new          34.1ms   37.9ms      26.4    1.43MB     1.69
#> 2 old          32.7ms   36.5ms      27.3    1.84MB     3.04

[EmilHvitfeldt]

Thank you loads!