mips64 ieee754 data structures

[Roland McGrath]

> > If we can replace 5 copies in 5 files with 2 copies in one file, I call
> > that an improvement.  It also leaves less for a new port to do, and
> > less to be done for a new configuration on an existing machine that
> > changes the `long double' size from what was used before.
> 
> No.  Exposing anything in headers which isn't true for a platform must
> be avoided.

Nonsense.  We have installed headers that test endianness and word size
even on platforms where it never varies, for example.  I don't think there
is anything wrong with that, and this is just the same.  We have always
balanced the clean minimalness of the installed header contents with the
cleanness of minimal code duplication in the source tree.