forder incorrect for na.last=TRUE and double column, descending character column

[MichaelChirico]

Found while working on #7837; Gemini created this minimal reprex:

DT = data.table(
  d1 = c(1, 10, 10, 10),
  c2 = c('a', 'b', 'c', NA)
)

DT[order(d1, -c2, na.last=TRUE)]
#       d1     c2
#    <num> <char>
# 1:     1      a
# 2:    10   <NA>
# 3:    10      c
# 4:    10      b

NA must be at the end; c.f.

# equiv. DT[base::order(d1, -xtfrm(c2), na.last=TRUE)]
DT[base::order(-d1, c2, na.last=TRUE, decreasing=TRUE)]
#       d1     c2
#    <num> <char>
# 1:     1      a
# 2:    10      c
# 3:    10      b
# 4:    10   <NA>

Note that {base} radix sort is not affected which hinted at this being a regression:

DT[base::order(d1, c2, na.last=TRUE, method='radix', decreasing=c(FALSE, TRUE))]
#       d1     c2
#    <num> <char>
# 1:     1      a
# 2:    10      c
# 3:    10      b
# 4:    10   <NA>

Gemini helped identify #3124 as the culprit by creating a branch off of 88439d9 (parent of #3124 == e59ba14), applying some minimal changes to get it compiling under r-devel as a first commit, then cherry-picking e59ba14 as a second commit. The example above WAI before e59ba14 is applied.