QBnd: sample the boundary texture with POINT filtering

[ggalgoczi]

cudaFilterModeLinear blends ~0.4% from the adjacent (categorically distinct) material/surface row on large (>5000-row) bnd tables, corrupting every per-photon material/surface lookup. Add a POINT-mode path selected by QBND_FILTER_POINT (run with QBND_FILTER_POINT=1).

[plexoos]

I agree that boundary_lookup likely needs a correction here. As described, cudaFilterModeLinear appears to introduce cross-row blending on sufficiently large boundary textures, and that is a correctness issue because the y axis indexes categorical material/surface rows.

A few follow-ups would make the change easier to review:

• Please expand the PR description in plain English. It would help to explain that the boundary table is packed into a 2D texture, x is wavelength, y is the material/surface row, why linear filtering was attractive originally, and why cross-row blending is a problem here.
• Please add a reproducer that demonstrates the impact at the physics level if possible, for example wavelength distribution, hit counts, or another observable using one of the existing test geometries. At minimum, a focused regression test showing the cross-row contamination on a tall bnd table would help.
• Please route the toggle through the existing Config machinery rather than introducing a raw env var as the primary interface. An env var override is fine, but the default should preserve the current behavior for now.
• If this remains configurable, the configuration should gate the full lookup strategy consistently. Right now the manual interpolation path changes in qbnd.h, while the texture filter mode only changes when the env var is set.

[ggalgoczi]

I agree that boundary_lookup likely needs a correction here. As described, cudaFilterModeLinear appears to introduce cross-row blending on sufficiently large boundary textures, and that is a correctness issue because the y axis indexes categorical material/surface rows.

A few follow-ups would make the change easier to review:

* Please expand the PR description in plain English. It would help to explain that the boundary table is packed into a 2D texture, `x` is wavelength, `y` is the material/surface row, why linear filtering was attractive originally, and why cross-row blending is a problem here.

* Please add a reproducer that demonstrates the impact at the physics level if possible, for example wavelength distribution, hit counts, or another observable using one of the existing test geometries. At minimum, a focused regression test showing the cross-row contamination on a tall `bnd` table would help.

* Please route the toggle through the existing `Config` machinery rather than introducing a raw env var as the primary interface. An env var override is fine, but the default should preserve the current behavior for now.

* If this remains configurable, the configuration should gate the full lookup strategy consistently. Right now the manual interpolation path changes in `qbnd.h`, while the texture filter mode only changes when the env var is set.

Thanks for the review. Pushed three follow-up commits addressing all four points.
PR description — I'll update it now. Short version:
The boundary table is a 2D CUDA texture: x = wavelength (continuous, 761 bins), y = material/surface property row (categorical — OMAT/OSUR/ISUR/IMAT × 2 property groups per boundary). With cudaFilterModeLinear the hardware does bilinear interpolation across both axes. Along x (wavelength) that's desirable. Along y it blends adjacent material rows — e.g. mixing glass refractive index with surface reflectivity. On small tables the y-coordinate lands close enough to texel centers that the contamination is negligible, but on tall tables (>500 rows, i.e. >60 boundaries) it becomes measurable.
Config toggle / consistent gating — Removed entirely. There's no valid use case for LINEAR on the y-axis — blending categorical rows is always wrong regardless of table size. The env var is gone; POINT filtering is now unconditional. This also eliminates the gating inconsistency (points 3 and 4).
Regression test — Added qudarap/tests/qbnd_boundary_lookup_test.cc, a CPU-only mock-texture test that verifies:

• Row isolation: adjacent rows with distinct values return exact per-row data, no blending
• Wavelength interpolation: manual two-bin lerp gives correct intermediate values
• Domain-edge clamping: wavelengths outside [nm0, nm_max] clamp to the boundary bin

Note on test scope: the mock texture system (s_mock_texture.h) does nearest-neighbor lookups, so it can't directly reproduce the LINEAR cross-row blending artifact — that requires a real CUDA texture on GPU hardware. The test pins the correctness of the POINT + manual-lerp code path.