Sogs updates

[slimbuck]

Fixes #7537

This PR updates SOGS runtime as follows:

• combine labels_u and labels_l data into single RG texture
• combine gaussian sh0 (color) and opacity into single texture
• implement more accurate quaternion storage in (2, 8, 8, 8) bits
• reorder data on GPU at load-time to improve runtime memory cache coherency

SOGS Data Comparison

Examples of SOGS compression. All scenes contain 3 bands of spherical harmonics.

Scene	Gaussians	PLY	Compressed PLY	SOGS
	4.3M	1GB	263MB	56MB
	60K	14.6MB	3.6MB	1.2MB
	720K	174MB	43MB	11MB

Notes

• rendering performance is still slower than compressed.ply (on my windows RTX3070 ti machine), but faster than uncompressed ply. haven't investigated why
• load time reordering isn't fast due to the cpu sort. 4M particles takes around 600ms (on laptop). we could perform this sort ahead of time and store the order in another texture, but for now we choose the load-time approach.
• the tooling for generating SOGS scenes is based on a fork of https://github.com/vincentwoo/sogs. we will make this public soon or merge our changes into the mainline, depending on @vincentwoo 's preference.
• creating SOGS scenes requires (currently) a CUDA-enabled python install. we hope to create a browser-based solution.

[Copilot]

Pull Request Overview

This PR updates the SOGS runtime to combine texture resources, improve quaternion storage precision, and reorder GPU data for better memory cache coherency.

• Combine separate label textures into a unified texture and merge Gaussian color and opacity data.
• Update quaternion unpacking and storage using a (2, 8, 8, 8) bit scheme.
• Introduce a GPU-based reordering step that computes Morton order indices to improve runtime memory performance.

Reviewed Changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/scene/shader-lib/chunks/gsplat/vert/gsplatSogsSH.js	Combined label uniforms into a single texture uniform.
src/scene/shader-lib/chunks/gsplat/vert/gsplatSogsData.js	Updated quaternion reading logic using a mode-based selection.
src/scene/shader-lib/chunks/gsplat/vert/gsplatSogsColor.js	Merged sh0 and opacity data into one texture and adjusted processing.
src/scene/shader-lib/chunks/gsplat/vert/gsplatData.js	Modified quaternion unpacking calls to support new bit storage order.
src/scene/shader-lib/chunks/gsplat/vert/gsplatCompressedData.js	Removed obsolete quaternion conversion function.
src/scene/shader-lib/chunks/gsplat/vert/gsplatCommon.js	Updated documentation and fixed minor comment typos.
src/scene/gsplat/gsplat-sogs.js	Removed redundant texture parameters and updated parameter passing.
src/scene/gsplat/gsplat-sogs-data.js	Added new GPU reordering logic based on Morton order for improved cache coherency.
src/scene/gsplat/gsplat-data.js	Clamped coordinate values to a maximum of 1023 to prevent potential overflows.
src/framework/parsers/sogs.js	Updated texture async reads and adjusted texture assignments for the new data layout.

[Copilot]

[nitpick] Consider adding an inline comment to explain the texture swap logic in this loop, clarifying how the chain of render targets is being updated for maintainability.

Suggested change

	// Swap targetTexture and sourceTexture to prepare for the next iteration.

[mvaligursky]

WGSL chunks need to be updated as well, otherwise this won't work on WebGPU

[mvaligursky]

Also, it'd be great to have an engine example using this, so that we can test both platforms easily.

[slimbuck]

There is some remaining follow up work:

• implement wgsl chunks
• get device recovery working
• add SOGS loader to engine examples

[Maksims]

If the rendering is done by reading from textures, then there might be a significant performance reduction when accessing texture texels at random positions. As it does not allow GPU to cache chunks of it, so the reads from it will require a slow memory trips.
This is especially strong on mobile.

[slimbuck]

@Maksims , yep, that's what this PR is addressing :)