segfault under MKL, ATLAS configurations on CRAN

[MichaelChirico]

Observed from CRAN:

https://www.stats.ox.ac.uk/pub/bdr/Rblas/ATLAS/data.table.out
https://www.stats.ox.ac.uk/pub/bdr/Rblas/MKL/data.table.out

Both have similar stack traces:

  Traceback:
   1: all.equal(ax, ay, ..., check.attributes = check.attributes)
   2: attr.all.equal(target, current, tolerance = tolerance, scale = scale,     ...)
   3: all.equal.numeric(x, y, check.names = !isTRUE(y))
   4: all.equal(x, y, check.names = !isTRUE(y))
   5: isTRUE(all.equal.result <- all.equal(x, y, check.names = !isTRUE(y)))
   6: test(1252 + test_no * 0.001, ans1[ans1 != 0], with(DT, eval(ll)))

ATLAS and MKL are different BLAS implementations:

https://www.stats.ox.ac.uk/pub/bdr/Rblas/README.txt

[MichaelChirico]

GitLab (test-lin-dev-gcc-strict-cran) is also finding this:

https://gitlab.com/Rdatatable/data.table/-/jobs/12544376798

[ben-schwen]

order(NA_character_, OTHER_STRING, method='radix', na.last=NA) errors now in R-devel as reported in r-devel mailing list

Since we explicitly test for these cases in 1252 we will also error. Could fix by switching to other method in order.