Avoid frollapply() stack overflows and protect limits

[DavisVaughan]

Closes #3993

The following three examples from #3993 now run without issue, no segfaults!

library(data.table)

xx <- frollapply(1, rep(1L, 1e5), identity)

yy <- frollapply(1, rep(1L, 1e6), identity)

zz <- frollapply(as.list(rep(1, 1e6)), 1, identity)

If someone else wants to take over adding tests (if you want them), please go ahead. I'm not familiar with how you all test the package.

[codecov]

Codecov Report

Merging #3994 into master will decrease coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff            @@
##           master   #3994      +/-   ##
=========================================
- Coverage    99.4%   99.4%   -0.01%     
=========================================
  Files          72      72              
  Lines       13652   13650       -2     
=========================================
- Hits        13571   13569       -2     
  Misses         81      81

Impacted Files	Coverage Δ
src/frollR.c	100% <100%> (ø)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update b8dd0db...9607681. Read the comment docs.

[jangorecki]

Thanks for fixes. The second commit is a nice change. I kept all objects pre-allocated to align to the code of other rolling functions, but because roll apply will not run in parallel we can safely override those variables inside the loop. Will add tests.

[jangorecki]

Tests were added but disabled by default as they require 1.6 GB, 16 GB and 16GB memory. There is no way to tests those failures in a lightweight way.