[VR] Quest 3 Stereo Rendering — Per-Eye Covariance, Shared Compute, Performance Fixes

[arghyasur1991]

Supersedes #173 with a cleaner, more correct implementation. Includes all fixes from the discussion there.

Summary

Adds single-pass instanced and multiview stereo rendering support for Quest 3 (and other VR headsets using URP), building on the approach from #173 with key correctness and performance improvements:

• Per-eye covariance projection: Passes explicit per-eye model-view and projection matrices to the compute shader via a StereoMatrices cbuffer. The original [VR] Single Pass Instanced/Multiview Support for Gaussian Splatting in URP #173 used UNITY_MATRIX_VP / _MatrixMV (center eye) for covariance math, causing visible artifacts ("thorny ferns") in VR. This fix computes correct 2D covariance per eye.
• Shared covariance/SH between eyes: Computes full covariance, SH evaluation, and color once for the left eye, reuses for the right — only the clip position is recomputed. Saves ~40% of compute shader ALU.
• Sort once, render twice: Sorting uses the center eye matrix (IPD offset is negligible for depth ordering), so per-eye sorting was removed — it produced identical results at 2x GPU cost.
• clip() instead of discard in the fragment shader for better Adreno TBDR performance.
• Dispatch fix: Uses m_SplatCount instead of m_GpuView.count for compute dispatch (view buffer is 2x for stereo).

Bug fixes (from #173 discussion)

• Editor stereo guard: !Application.isEditor prevents false isStereo detection with MockHMD/OpenXR in editor play mode (fixes #173 comment — right eye rendering black)
• Tex2DArray dimension check: Prevents stereo path for OVROverlayCanvas and other stereo-enabled but 2D-target cameras
• RT dimension mismatch: Uses cameraTargetDescriptor instead of XRSettings.eyeTextureDesc for RT creation, fixing depth/color surface dimension errors on Quest
• Sorter validity guard: m_Sorter.Valid check before dispatch for GPUs lacking wave intrinsics

Stereo rendering approach

The CSCalcViewData compute kernel now:

1. Checks _IsStereo uniform
2. If stereo: calls CalculateEyeViewData() with left-eye matrices (full covariance + SH), copies result for right eye with only the clip position recomputed from right-eye VP matrix
3. Writes to _SplatViewData[idx * 2] (left) and _SplatViewData[idx * 2 + 1] (right)
4. RenderGaussianSplats.shader indexes into view data via instID * 2 + eyeIndex

The URP feature detects stereo via XRSettings + Tex2DArray check, uses PrepareSplats() (sort + calc view data once) then RenderPreparedSplats() per eye, compositing through a stereo tex2darray composite shader.

What's NOT in this PR

• HDRP stereo support
• Debug point/box shader stereo variants
• unity_StereoEyeIndex usage (doesn't work with DrawProcedural on Quest 3 — manual 2-draw-call workaround used instead, see TODO in code)

Tested on

• Quest 3 (Multiview mode) — correct stereo rendering, stable frame rate
• Desktop non-VR — no regressions, single-eye path unchanged

[arghyasur1991]

not raising PR yet, but there are more performance optimizations present in https://github.com/arghyasur1991/UnityGaussianSplatting/tree/feature/quest-stereo-perf and my fork's main branch which allows rendering gaussian splats count as high as 300k at 16-18FPS in quest 3 without any jittering/visible frame drops (even from closeups).
That branch includes high impact changes like -

1. Early out culling based on opacity/scale and frustum - Medium impact without perceivable quality loss
2. Reduced-resolution rendering with bilinear upscale - High impact without much perceivable quality loss
3. Lower RT precision - High impact without perceivable quality loss