This isn't the _right_ fix, which would involve using collations
correctly and working out what language we wanted to sort using. But
assuming most people are using Excel in English, this fixes the most
common case where sort does not match; apostrophes in names.
Just removing the character does not lead to a consistent sort, however,
so it may be worth reviewing that if people notice.

This goes back to the in-place approach from the original PR, since
comparator chaining like this is nicer.

Neaten the tests, add more test cases for the apostrophe cases.

This still won't help with names identical except in an apostrophe in
the first name, but that is both much less likely and would make the
sorting still more expensive to compute (because the comparators are
nested in reverse order).