Tests robust to locale sorting

[MichaelChirico]

Closes #3502.

@minemR it would be great if you could confirm this works on Windows as well, but I don't have any reason to believe there's an issue there.

[tdhock]

I did a PR to fix some tests on a machine where I run a French locale, #4719
The approach I used there was not setting the locale, and instead making the expected values depend on the locale too.
This approach in this PR is setting the locale to C for the duration of the test, then setting back to original locale.
I don't have any problem with this approach, just wanted to mention the difference.

[MichaelChirico]

I mentioned an approach like that in the linked issue: #3502 (comment)

I might be missing something simpler, but the approach I was seeing to make the test reference depend on the locale was pretty abstruse, hence going with temporarily editing the locale.

IIUC, in #4719, data.table and the reference value both depend on the underlying locale setting, which may be why it's easier to avoid temporary settings in that case. Here, data.table's value is completely independent of the setting.

Actually, this discussion is reminding me that, because base R uses data.table's own code for method="radix", it's also independent of locale, so possibly that will work instead (though not for factor leveling, easily). However, that's only available in R>=3.3.0. I'll add a TODO to revisit that once we have a high enough R dependency.