Add hardware occlusion culling for camera-view passes

[tomconder]

Summary

• Add per-object GL_ANY_SAMPLES_PASSED occlusion queries, each testing an object's world AABB against the depth prepass. Objects with no visible pixels are skipped in the depth prepass and opaque pass.
• Extend the same technique to the shadow pass: a second, independent OcclusionCuller instance tests each object's AABB against the shadow map's own depth (light view/projection, not the camera's), alongside the existing light-frustum mask. An object can be light-occluded (hidden behind a closer shadow caster) without being camera-occluded, or vice versa, so this is a separate result, not a reuse of the camera-view one.
• Results lag one frame in both passes: pollResults() picks up last frame's answers before deciding what to draw, query() runs after this frame's real depth is rasterized so the answer is ready for next frame. For the shadow pass, the query runs right after the shadow depth renders, while its FBO is still bound.
• An object whose AABB contains the query's eye position skips the query and stays visible. From inside, a box only shows its far face (the near face is behind the eye, clipped away), which is always farther than the object's own nearby geometry already in the depth buffer — so the query would fail against itself every frame. Sponza is exactly this case for the camera: its AABB is the whole level, and the camera never leaves it. ShadowMap now exposes the light's view eye position so the shadow-pass query can apply the same guard. The orthographic light projection doesn't clip this way, so there the guard is just a harmless no-op.
• Query draws also need GL_CULL_FACE off: the renderer's global back-face culling would cull every face of a box the eye is standing inside, leaving nothing for the query to rasterize.
• New OcclusionCuller class in sponge/src/platform/opengl/scene/, following the same one-class-per-GPU-resource pattern as ShadowMap/ClusteredLights/Cube. Its query() takes a generic eyePos (not cameraPos) since it's shared by both the camera and the light. Shared the unit-cube geometry between Cube and the new query proxy via unitcube.hpp.
• Debug UI: "Meshes" now plots a 100-frame rolling graph of frustum-visible mesh count (ImGui::PlotLines, overlay shows current/total). "Objects" stays a plain X / Y occlusion-visible text row.

Benchmark

Measured on the current maze.yaml scene (release build, ENABLE_IMGUI=ON), standing in the main hall:

Metric	Value
Objects	3 / 3 occlusion-visible
Meshes (frustum)	66 / 105 visible
Cull / submit	2-3 us / 258-266 us
Frame	~3.0-3.1 ms (320-330 FPS)

This scene has only 3 top-level objects: Sponza (one object holding all 105 meshes), the helmet, and the debug cube. Sponza's own AABB is the whole level, so per the eye-inside rule above it's permanently exempt from testing against both the camera and the light — meaning the 105 meshes that dominate this scene's cost are never reachable by either pass. The other two objects are both near the spawn point and stay in view/lit together. Object-granularity occlusion culling is a no-op on this scene today, camera and shadow alike, same conclusion as #499 for the light-frustum shadow cull: the win shows up once the scene has separate objects that can actually occlude each other (a real maze layout with separate wall/room objects), not on a single-mesh backdrop.

• Query overhead itself is invisible at this object count: submit time (which includes the skip check) is unchanged within noise from a plain frustum-only baseline.

Test plan

• cmake --build out/build/ci-windows-release --target game succeeds
• pre-commit passes
• Launched the built game, confirmed Sponza, the helmet, and the debug cube all render correctly and stay correct after moving the camera
• Confirmed shadows still render correctly with shadow-pass occlusion culling enabled
• Confirmed no GL_INVALID_OPERATION or other driver errors over an extended run (caught and fixed one: polling a query before its first glBeginQuery/glEndQuery)
• Confirmed the Meshes graph renders and scrolls correctly in the debug UI