27_io/istream_extractor_arith.cc fixup

[Loren James Rittle]

>> test case.  It is still failing, at least HP and FreeBSD, due to
>> undetected overflow on double to float cast in a non-C99 code path.

>Huh. Bummer. Is there a way to fix this up?

Yes and no.  There is no obvious fix portable to all failing machines.

I have a patch to fix BSD and Solaris but David and I went a few
rounds trying to get the technique to support HP without success.

The best fix would be to add some stuff to libmath and get rid of all
macro selected regions out of the mainline of code, but I haven't had
the time to figure out how to do it yet.