Just use UTF8.

Turns out a fair amount of the problems were with natsort itself.

FreeBSD seems to throw an OSError when locale.strxfrm is given 'Å',
which is surprising behavior. Well, maybe not, considering how many bugs
I have found with FreeBSD's implementation of locale over the course of
natsort development.

Anyway, we just ignore any input that causes locale.strxfrm to barf in
our tests.

The previous code change to make NaN and None ordering consistent made
it so that NANLAST did not put NaN last. Oops.

It also had made it so that NaN wasn't first for NANFIRST. Oops.

This will sort the collection as strings before sorting with the natsort
algorithm. This ensures that strings that are different but represent
the same numerical value get sorted independent of input order.

Internally, these may be translated to the same value, so they will be
output in the same order they were input, which could lead to suprise.

This commit ensures the order is always consistent.

The new version can convert the entire mapping internally, so we use
that if possible. A new wrapper for this new function is introduced to
make the calling code consistent.