Gsplat sorting update

[slimbuck]

This PR updates the gsplat sorter when rendering compressed data:

• send the compressed chunk min/max bound to sort worker
• use chunk bounds to calculate a per-frame histogram of chunks based on distance-to-camera
• use the histogram to allocate counting sort key bits

This results in much more stable rendering of our most challenging (especially MCMC) scenes, but also (unexpectedly) better rendering on many others.

The uncompressed sort is sped up slightly, but otherwise remains unchanged.

Cost:

• cpu sort seems to be around 20% slower
• additional memory storage requires 6 x float32 per 256 gaussians (which practically very small)

Before:

Screen.Recording.2025-03-24.at.15.25.27.mov

After:

Screen.Recording.2025-03-24.at.15.26.33.mov

Before:

Screen.Recording.2025-03-24.at.15.33.47.mov

After:

Screen.Recording.2025-03-24.at.15.34.45.mov

Before:

Screen.Recording.2025-03-24.at.15.35.59.mov

After:

Screen.Recording.2025-03-24.at.15.36.54.mov

Notes

• default histogram size is arbitrarily set to 32. More buckets just seem to result in more stable sorting, but cost more CPU time.
• would be nice to support this approach for uncompressed format, but without chunks it would probably run twice as slowly

Fixes #6817