aephysics.manifold: contact manifolds for spheres, capsules and hulls

[nicolas-maman]

The sixth layer: Box3D's manifold.c + convex_manifold.c in Aether, as aephysics.manifold.

• Hull–hull: the separating axis test (A's faces, B's faces, edge pairs whose Gauss-map arcs cross — the reference's scalar path exactly), the SAT cache (answers when the last axis still separates or still builds the same contact), incident face from the closest vertex, Sutherland-Hodgman clipping to the reference face with feature pairs on every point, reduction to four points with the reference's tie-break, face contact by preference / edge contact when clearly better.
• Sphere–sphere, capsule–sphere, hull–sphere, capsule–capsule, hull–capsule (GJK shallow, face and edge axes deep).
• test_manifold.ae: 43,090 checks — test_sat.c's brute-force oracle over 7,000 random box pairs (offset hulls included) and the hull/sphere/capsule parts of test_manifold.c. The triangle parts wait for the triangle manifold.
• bench/manifold.ae vs bench/manifold_box3d.c: identical manifolds (point counts, cache hits, separation sums equal on every phase). Warm-cache hull–hull at parity (3.7 vs 3.6 ms / 20k); cold SAT 1.6× (theirs is SIMD); capsule and sphere 1.5×.
• Found typeck/codegen: an array literal of float expressions (0.18 * math.PI) is typed as int and truncates every element aether#2119 on the way: an array literal of float expressions (0.18 * math.PI) is silently typed as int.

Differences: clip buffers are module scratch (one collision at a time until the parallel layer); the 4-point reduction is the reference's default variant.