Path Tracer Light Cut Trees and Alias table

[karimsayedre]

Description

Adds many-light next-event estimation to the path tracer: lights are picked either by stochastic lightcuts-style tree or a flat alias table, with MIS-exact probabilitiesn both directions, and the chosen light's solid angle is then sampled with theamplers from #1058 (new-sampler-concepts) and builds upon it. Also adds the GUI and camera work that makeshe path tracer usable and navigatable.

Light tree (sampling/stochastic_lightcut_tree.hlsl)

• 4-ary heap (children at 4i+1..4i+4, leaves padded to power-of-4) with a canonical 32 B CWBVH-4 packed wide node: fp32 origin, fp16 parent power, one shared biased exponent, 4-bit leaf mask, and per-child 4-bit quantized bboxes (floored min / ceiled max, so decoded boxes always contain the true ones) + 8-bit relative power. One encode/decode contract shared by the CPU builder (CLightTree.cpp) and the GPU accessor.
• Receiver-adaptive descent with compile-time weight modes (0: full geometric, 1: power-only, 2: uniform, 3: power × orientation, the default; distance importance belongs in the RIS target, not the proposal denominator). Fully-facing clusters take a transcendental-free fast path for the cosine cone bound. MAX_RATIO early stop (Estevez-Kulla "no clear winner"): descent hands the remaining pick to a per-subtree power-weighted alias table in O(1), via the SubtreeAliasAccessor concept. Fix included: the old default τ=0.2 was unsatisfiable (max of 4 non-negative weights ≥ wSum/4), so the stop never fired; default is now 0.5 and the >¼ feasibility bound is documented.

Alias table (sampling/alias_table.hlsl)

• Used both as the flat NEE selector and for the tree's per-subtree tables. Packed layout A (4 B word + separate pdf[], 8 B/sample).

Renderer (ex40)

• NEE selector switchable between light tree and flat alias table; beauty shader has three MIS modes (NEE_ONLY, BxDF_ONLY, BOTH) for unbiased A/B anchoring.
• NEE_PROPOSAL_PROBE debug mode visualizes first-bounce contributing samples;
• Very experimental debug shader shows the quantized cluster AABBs.

Testing

• examples_tests 37: CPU builder tests (8 weight distributions), GPU correctness testers for alias/CDF/tree (forward/backward pdf + weight consistency), and the discrete-sampler benchmark sweep (N = 2 → 33.5M, ps/sample via GPU timestamps using the framework from Add structured stats/IRs to pipeline executable info #1070(IBenchmark)), with serialized-RNG loops to prevent unrealistic compiler unrolling/pipelining.
• ex40 benchmark reports per-row details including Beauty images, usable for bias and FLIP checks; fixed the spp-cap bug that produced no-op frames in cheap rows.