Port better to x86 -fexcess-precision=fast

Port better to x86 -fexcess-precision=fast

Problem reported by Eli Zaretskii in:
https://lists.gnu.org/r/emacs-devel/2018-08/msg00380.html
* src/data.c (arithcompare): Work around incompatibility
between gcc -fexcess-precision=fast and the C standard on x86,
by capturing the results of floating-point comparisons before
the excess precision spontaneously decays.  Although this fix
might not work in general, it does work here and is probably
good enough for the platforms we care about.