The POSIX spec says that rint() rounds halfway cases to nearest even.
Our substitute implementation failed to do that, rather rounding halfway
cases away from zero; and it also got some other cases (such as minus
zero) wrong.  This led to observable cross-platform differences, as
reported in bug #12885 from Rich Schaaf; in particular, casting from
float to int didn't honor round-to-nearest-even on builds using rint.c.

Implement something that attempts to cover all cases per spec, and add
some simple regression tests so that we'll notice if any platforms still
get this wrong.

Although this is a bug fix, no back-patch, as a behavioral change in
the back branches was agreed not to be a good idea.

Pedro Gimeno Fortea, reviewed by Michael Paquier and myself