Check all clip vertices in the spherical Sutherland-Hodgman containment fallback

[asinghvi17]

Fixes #467.

The defect

When the spherical _intersection_sutherland_hodgman clips the subject down to
nothing, it falls back to asking whether the clip polygon was inside the subject
all along. That check only tested clip_points[1]:

if !isempty(clip_points) && _point_in_convex_spherical_polygon(clip_points[1], original_subject)

_point_in_convex_spherical_polygon uses spherical_orient >= 0, so the boundary
counts as inside. Any subject that merely touches the clip ring's first vertex
passed the test and was handed the clip polygon's entire area.

Instrumenting the three grazing tiles from the issue's MWE shows it directly —
for each one, output = 0 pts and the per-clip-vertex containment vector is
[1, 0, 0, 0, 0]. One vertex in, four out, whole clip returned. Summed over the
three disjoint tiles that is 3x the clip's own area; in the reported polar case
it is 145x, where every degenerate quad of a 216-column mesh shares the pole.

The fix

Require every clip vertex to be inside:

if !isempty(clip_points) &&
   all(p -> _point_in_convex_spherical_polygon(p, original_subject), clip_points)

Two things I verified before landing this, both of which say it is free:

• The branch is unreachable for valid input. A sweep of 159600 convex-convex
  pairs hits the empty-output branch 158732 times and finds clip_points[1]
  inside the subject zero times. Re-running the sweep after the change gives an
  identical result, so this is a no-op in contract.
• Genuine containment never reaches it anyway. Ordinary Sutherland-Hodgman
  already returns the right answer when the clip is inside the subject — for a
  small quad inside a big one the output is 4 points, not empty. The branch's
  stated purpose is handled upstream.

It stays sound out of contract too: _point_in_convex_spherical_polygon on a
non-convex subject under-approximates (it answers about the half-space
intersection, a subset), so it cannot produce a false positive.

The cost is O(m) point tests instead of 1, on a branch that is already the slow
path. Only the spherical path has this fallback; the planar path has none.

Test

Adds "Grazing tiles are not handed the whole clip". Written first: it failed 4/4,
at exactly 1.0x the clip area per tile and 3.0x summed, then passed.

One caveat worth flagging for review: as noted in #467, reaching the fallback at
all takes a non-convex clip, so the fixture is out-of-contract input — with a
convex clip a grazing subject survives every half-space test and leaves a 1-2
point output that the degenerate branch already zeroes. The test is therefore
framed as graceful degradation: the answer may be wrong, but it must be wrong
towards zero, never towards a whole fabricated cell. That is a realistic concern
rather than a contrived one, since a near-convex cell can go non-convex
numerically (see #466). If an out-of-contract fixture is unwanted in the suite,
the fix stands on the in-contract sweep alone and the test can be dropped.

Verification

Full suite passes (Testing GeometryOps tests passed); Sutherland-Hodgman 69/69.
The issue's MWE now returns exactly 0.0 for all three tiles.

🤖 Generated with Claude Code