1.0.3.1: fix behaviour of >= and <= with NaNs

1.0.3.1: fix behaviour of >= and <= with NaNs

 * Problem: (>= (/ 0.0 0.0) 1.0) evaluates to true.

   Move inversion from >= to < and from <= to > from the source
   transformations to new deftransforms, and make it conditional on
   the derived type, avoiding the inversion for potential floats.

   This fixes the NaN issues with >= and <=, but exposes gaps in our
   transformations for =, causing a constraint propagation test to
   fail...

 * Tweak the deftransform for = so that another transformation can
   co-exist with it.

   Write INTERVAL-= and INTERVAL-/=, and deftransforms for = and /=
   based on them.

   This fixes the constant propagation issue, except for the bug it
   exposes elsewhere...

 * INTERVAL-INTERSECTION/DIFFERENCE returns bogus intersections -- fix
   it.

   ...and all is well.

 * Tests.

 * NUMERIC-TYPE-OR-LOSE is unused, deleted.

 * Also fix tests/stream.impure.lisp for UTF-8 environments.