Real.maxFinite bug

[Matthew Fluet]

> > I agree: all these constants should just be specified as constants given the
> > knowledge of IEEE reals, not computed.
> > Mind you, the computation is useful for detecting problems, but I would move it
> > to the regression tests.
> 
> OK.  We're settled.  Matthew, can this go in your queue?

Yeah.  I guess I'll just bounce them through PackReal.* functions, but
since those are implemented as _ffi's, then they won't be real compile
time constants.  (Also, I'll have to see about getting the ordering right;
this might require moving pack-real.sml before real.sml in build-basis,
but I think that's o.k., because there aren't any direct dependencies
between the two.)