WebGPU: add LightProbeGrid example

[querielo]

Summary

This PR adds WebGPU light probe grid support and expands the examples to cover single-volume, multi-volume, and Sponza scenes. The WebGPU path remains separate from the existing WebGL LightProbeGrid addon.

Included

• add webgpu_lightprobes.html
• add webgpu_lightprobes_complex.html
• add webgpu_lightprobes_sponza.html
• add screenshots for the new examples
• register the new examples in examples/files.json
• add LightProbeGridGPU.js so probe grids can be added to the scene and baked directly

Notes

• WebGPU path is kept separate from the existing WebGL addon
• examples can use scene.add( probes ) and await probes.bake( renderer, scene, options )
• multi-volume sampling automatically uses object-origin bounds selection, matching the existing WebGL behavior
• single-volume sampling avoids bounds rejection, matching the existing WebGL behavior

Demo

• GitHack single volume: https://raw.githack.com/querielo/three.js/kirill/webgpu-lightprobes/examples/webgpu_lightprobes.html
• GitHack multi volume: https://raw.githack.com/querielo/three.js/kirill/webgpu-lightprobes/examples/webgpu_lightprobes_complex.html
• GitHack Sponza: https://raw.githack.com/querielo/three.js/kirill/webgpu-lightprobes/examples/webgpu_lightprobes_sponza.html

[querielo]

@sunag @Mugen87 Could you help me?

% npm run test-e2e-webgpu webgpu_lightprobes_sponza

> three@0.184.0 test-e2e-webgpu
> node test/e2e/puppeteer.js --webgpu webgpu_lightprobes_sponza

Diff 0.0% in file: webgpu_lightprobes_sponza
TEST PASSED! 1 screenshots rendered correctly.
Closing...


% npm run test-e2e webgpu_lightprobes_sponza

> three@0.184.0 test-e2e
> node test/e2e/puppeteer.js webgpu_lightprobes_sponza

Diff 0.0% in file: webgpu_lightprobes_sponza
TEST PASSED! 1 screenshots rendered correctly.
Closing...

[querielo]

@sunag @Mugen87 @mrdoob forceWebGL: true and forceWebGL: false produces different images and the difference is not LightProbeGrid.

Isolation results:

full WebGL fallback (forceWebGL) vs native WebGPU: 0.199% diff
GI disabled:                                       0.251% diff
shadows disabled:                                  0.126% diff
GI + shadows + light + sky disabled:               0.803% diff
model hidden too, flat background only:            0.000% diff

So source is mesh edge rasterization / MSAA coverage between WebGLBackend and WebGPUBackend. Bright high-contrast geometry edges differ slightly. That is why JPEGs differ by ~0.2-0.3%, even though visually same.

I'll try to disable MSAA for now to get the tests passing.

[querielo]

@sunag I’ve just updated LightProbeGridGPU, it should be much easier to use now. What device are you using to generate the previews?

[sunag]

@sunag I’ve just updated LightProbeGridGPU, it should be much easier to use now. What device are you using to generate the previews?

I’m taking the screenshot directly from the zip that the E2E testing is generating internally in a more private way (as colaborator), because the Windows screenshots I’ve been taking recently have been failing by minimal percentages.

[Mugen87]

Actually, we wanted to wait with a port until LightProbeGrid gets more stable. It's up to @mrdoob to decide an appropriate point in time.

IMO, at least #33658 should be done before focusing on a port.

[Mugen87]

Closing for now. We make a fresh PR when #33658 is done.