fix: use logarithmic depth quantization in compute GSplat bitonic sort

[mvaligursky]

Fix sort-instability flickering in the compute GSplat renderer by switching the bitonic sort's depth quantization from linear to logarithmic.

Changes:

• Replace linear depth quantization with logarithmic quantization in the per-tile bitonic sort, matching the approach used by the bucket sort for large tiles
• This gives proportionally more depth resolution to nearby splats and reduces depth-level collisions in tiles where perspective causes many splats to occupy a narrow depth range
• Eliminates visible frame-to-frame flickering in depth-complex areas (e.g., foliage) caused by non-deterministic atomicAdd ordering in the scatter pass producing unstable sort results for equal-depth splats

Performance:

• Negligible cost: one additional log() per splat in the sort setup phase (GPU SFU instruction), while the dominant Phase 4 bitonic comparison network is unchanged