[Enhancement] forder na.last=NA on characters

[arunsrinivasan]

forder on char with sort=FALSE is great (for speedup). And na.last=FALSE and na.last=TRUE need not hold much weight (since sort=FALSE).

However, na.last=NA implies omitting NA rows, sorting has no say here, I think. i.e., irrespective of whether sort=FALSE or TRUE, na.last=NA should not return indices of NAs in order indices.

Right now:

x = c("b", "b",  NA, "a", "b", "a")
data.table:::forderv(x, sort=FALSE, na.last=NA, retGrp=TRUE) 
[1] 1 2 5 3 4 6
attr(,"starts")
[1] 1 4 5
attr(,"maxgrpn")
[1] 3

but it should be:

[1] 1 2 5 0 4 6
attr(,"starts")
[1] 1 4 5
attr(,"maxgrpn")
[1] 3

This is useful in at least one scenario. Speeding up uniqueN when na.rm=TRUE. See #1455.

[arunsrinivasan]

Hm, doesn't seem to speedup any better than sort=TRUE. In fact, sort=FALSE is slower (on 100,000 unique values). Sticking to current approach.