Implement as_disjoint_union for unions of regions

[Chessing234]

What happens

Region.as_disjoint_union is how a region gets rewritten into the form the rest
of the Python code can work with: a list of convex polytopes that cover it
without overlapping. For Region_Type.UNION it did not exist —

if self.region_type == Region_Type.UNION:
    raise NotImplementedError(self.region_type) # TODO: implement this

so a region built with Region.union could not be converted at all, and neither
could anything containing one.

Fix

Each child's pieces are trimmed against the pieces already accepted, with
Polytope.set_minus, which splits A \ B into pairwise disjoint parts. The
accumulated list therefore stays disjoint at every step and its union is
unchanged, which is exactly the invariant the caller needs.

Subtracting a bounded piece from another leaves unbounded parts routinely, and
those reach Polytope.try_union through the simplification helpers, where they
hit the finite-polytope guard:

>>> Polytope.try_union([Polytope([[0, 1, 0]]), Polytope.rect((0, 1), (0, 1))])
ValueError: Polytope.try_union only supports finite polytopes (an input has extreme rays)

try_union answers "do these have a convex union?", and every caller already
treats None as "no, keep them separate". Raising turns that routine answer
into a crash of the whole pipeline, so the guard now returns None. The
vertex-only feasibility checks still cannot see extreme rays, so the check
itself stays.

Verification

tests/test_region.py gains a case in the style of the ones next to it: 50
trials of three random rectangles, with real overlaps.

• the disjoint pieces contain exactly the points the union contains
• no two pieces share an interior point (p.intersect(q).is_empty(include_boundary=False))

Ran separately over random 3-dimensional boxes as well — 2400 sample points
across 6 configurations, no point covered zero times that should be covered, and
none covered twice.

tests/test_region.py passes. Note that tests/test_all.py still fails on
test_affine2's import of large_values, which is the breakage #200 fixes;
this branch does not touch it.