Collect output in an accumulating stack

[hadley]

I think this slightly simplifies the existing code, and as you'll seen in upcoming PRs, allows us to tease apart concerns to yield code that's a little easier to understand.

---

You might be surprised to see me repeatedly growing a list, but this is quite performant in modern versions of R:

f <- function(n) {
  output <- watchout()
  for (i in 1:n) output$push(1:10)
  output$get()
}
bench::mark(
  f(10),
  f(100),
  f(1000),
  f(10000),
  check = FALSE
)[1:3]
#> # A tibble: 4 × 3
#> expression      min   median
#> <bch:expr> <bch:tm> <bch:tm>
#> 1 f(10)      286.02µs 368.92µs
#> 2 f(100)     352.27µs 442.66µs
#> 3 f(1000)      1.03ms    1.2ms
#> 4 f(10000)     7.79ms   8.49ms

(And for older versions of R, I don't think it would the typical length of output is going to cause a performance bottleneck)

[hadley]

This error handling is a little inelegant, but I'm pretty sure it'll go away in a future PR.

[cderv]

This makes sense. I did not know the performance had improved. It is cleaner to use this push() logic at least for reading code and understanding what is happening!