Fix division-by-zero in null-scattering PDF computation for participating media

[DeepReinforce]

Summary

Fix division-by-zero bugs in the null-scattering path of participating media sampling. When the null-scattering PDF (T_maj[0] * sigma_n[0] or T_maj[0] * sigma_maj[0]) evaluates to zero, the original code performs division before checking the zero condition, producing NaN/Inf values that propagate through beta, r_u, r_l, and T_ray, corrupting the entire path contribution and potentially causing rendering artifacts.

This PR moves the zero-check before the division and immediately returns false to terminate the path, avoiding any undefined arithmetic.

Changes

CPU integrators (src/pbrt/cpu/integrators.cpp)

• VolPathIntegrator::Li (null-scattering callback): Check pdf == 0 before dividing beta, r_u, r_l by pdf. Set beta to zero and return false early.
• RandomWalk (null-scattering callback): Same fix — check pdf == 0 first, set beta to zero and return false instead of conditionally skipping only the beta update.

Wavefront media sampling (src/pbrt/wavefront/media.cpp)

• WavefrontPathIntegrator::SampleMediumInteraction (null-scattering callback): Check pr == 0 before dividing beta, r_u, r_l by pr. Set beta to zero and return false early.

Wavefront transmittance (src/pbrt/wavefront/intersect.h)

• TraceTransmittance (ratio-tracking null-scattering): Add pr == 0 guard before dividing T_ray, r_l, r_u by pr. Set T_ray to zero and return false early.

Motivation

In scenes with participating media, it is possible for the null-scattering probability T_maj * sigma_n (or T_maj * sigma_maj) to evaluate to exactly zero. The original code either:

1. Divides first, then checks for zero (CPU path, wavefront media), or
2. Has no zero-check at all (wavefront transmittance).

Both cases lead to NaN/Inf propagation, which can cause visible fireflies, black pixels, or assertion failures depending on the scene configuration and medium parameters.