Rewrite reduce() and family in C

[ErdaradunGaztea]

Closes #1168 and closes #1243. Your acceptation of #1169 prompted me to tackle a bigger task 🦔

This PR reimplements reduce() (and reduce2(), accumulate(), and accumulate2() as well) in C. The idea came from Hadley's comment on #1168 stating that it'd be hard to performantly add progress bars to reduce() as they are written in C, so I rewrote the function in C instead. It is more performant too, getting reduce() to be on par with base Reduce():

x <- as.list(1:20000)
# Without .init or .progress
bench::mark(
  purrr::reduce_old(x, sum),
  purrr::reduce(x, sum),
  Reduce(sum, x),
  min_time = 1
)
#> # A tibble: 3 × 6
#>   expression                     min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 purrr::reduce_old(x, sum)   15.1ms  15.37ms      61.9     762KB     81.8
#> 2 purrr::reduce(x, sum)       8.08ms   8.35ms     118.      170KB     73.6
#> 3 Reduce(sum, x)              7.72ms   8.05ms     124.      349KB     51.6

^{Created on 2026-01-18 with reprex v2.1.1}

accumulate() isn't quite as fast as in base R, but there's improvement nonetheless:

x <- as.list(1:20000)
bench::mark(
  purrr::accumulate_old(x, sum, .init = 0),
  purrr::accumulate(x, sum, .init = 0),
  Reduce(sum, x, init = 0, accumulate = TRUE),
  min_time = 1
)
#> # A tibble: 2 × 6
#>   expression                            min  median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                        <bch:t> <bch:t>     <dbl> <bch:byt>    <dbl>
#> 1 purrr::accumulate_old(x, sum, .i… 26.2ms  27.3ms       36.8    1.45MB    178. 
#> 2 purrr::accumulate(x, sum, .init … 12.92ms 14.14ms      69.7    1.47MB    103. 
#> 3 Reduce(sum, x, init = 0, accumul…  9.33ms  9.86ms     100.   505.53KB     73.1

^{Created on 2026-01-28 with reprex v2.1.1}

I tried not to change any behavior (except for #1243) and hopefully succeeded, though that meant adding a lot of tests in the end. That was also the reason I didn't tackle #1245, as it'd change non-erroring behavior.

As for the code, well... It's not pretty, but I couldn't think of any approach that would look cleaner with interacting .init, .dir, and done(), and trying to make one shared implementation for all four functions with only some differences in certain points of the algorithm. If you prefer, I can separate C loops for reduce() and accumulate() family, that would probably be the most influential one. If I were to say, it's probably 15-20% more complex than the current R implementation (numbers completely made-up).

Finally, the C implementation should make it very easy to add .dir parameter to reduce2() and accumulate2(), as well as make accumulate2() return a named vector if input is named, like accumulate() does. Then again, I didn't want to add more changes than necessary and I can always add it later if you so wish.

No worries, I'm quickly running out of purrr functions to optimize :)

[ErdaradunGaztea]

Note that my last commit was necessitated by the latest vctrs (I had v0.6.5 installed) release adding extra information to the errors. If you'd rather snatch these updated snaps out of this PR because they do not fit the theme here - you're welcome.