Make hidden geometry sampling independent of scene shape ordering

[matla91]

The hidden-geometry sampling distribution is currently built in the order returned by scene.shapes(). The sampled distribution index is then used directly with scene.shapes_dr(). If the shape order changes, the same HGS sample can therefore select a different physical shape.

This change sorts the HGS categories by shape.id() when all scene shapes have unique, non-empty IDs. It stores the corresponding original shape indices and maps the sampled category back to the correct index before gathering from scene.shapes_dr(). Scenes without unique, non-empty IDs keep the previous behavior.

Regression test

I added a test that builds the same scene twice, with the two hidden shapes declared in opposite orders, and evaluates the same HGS sample in both scenes.

On the current main branch:

shape_a, shape_b -> [-2.0, 0.0, 1.0]
shape_b, shape_a -> [ 1.25, 0.0, 1.0]

With this patch:

shape_a, shape_b -> [-2.0, 0.0, 1.0]
shape_b, shape_a -> [-2.0, 0.0, 1.0]

The sampling PDF is unchanged.

The regression test passes with llvm_ad_rgb and cuda_ad_rgb. I also ran it in 10 fresh processes on each backend; all runs passed.

pytest still reports the existing failure in tests/integration/test_nlos.py, where the test calls prepare_transient(). That method is not present on the current TransientNLOSPath, and I reproduced the same failure on main before making this change.