Collision cleanups: test-only code, serialization helpers, stub aliases, tests #3826

Small cleanups left from the collision reviews (#3789, #3791), none of them affecting behaviour:

• Production code used only by tests:
  • cuda::collisionSortWorkspaceBytes (collision.cpp:12-15);
  • collision_byte_budget and the byte_budget parameters of encodeCollision/decodeCollision (project_document.cpp:2104, 2213-2217, collision_serialization.hpp:16-20, 49-50).
• Unreachable branches: max_size() and checked_product in collisionOutputBytes, behind the 64M check.
• collision_serialization.hpp lives in core but only project_document.cpp uses it, and it hand-rolls little-endian put/get. Moving it to io/project lets it use chapter_binary_utils.hpp (read_u32, write_u32, read_f32), about 35 lines less.
• captureCollisionSources always returns one source (scene.cpp:4348), and the prepare kernel's offset is always 0.
• The lf aliases of the collision functions (module.cpp:3372-3373) appear twice in the stubs.
• Tests:
  • CaptureSuccessAndExceptionFence… can be about 25 lines instead of 77;
  • OutputLayout…/BoundsOutput… test arithmetic and CUB behaviour;
  • test_collision_io.py only checks that the functions exist and that a missing file fails, where a small round trip would test more;
  • a loop asserts that the argument library rejects unknown flags.