LightsNode: Fix castShadow regression.

[Mugen87]

Related issue: https://discourse.threejs.org/t/webgpu-shadow-tests-and-findings/82561

Description

Toggling light.castShadow with WebGPURenderer has currently no effect.

The bit that checks the light state per frame is LightsNode.customCacheKey(). In this method we have to implement all checks that detect state changes of lights.

[github-actions]

📦 Bundle size

Full ESM build, minified and gzipped.

	Before	After	Diff
WebGL	337.26 78.64	337.26 78.64	+0 B +0 B
WebGPU	549.58 152.34	549.57 152.34	-16 B +1 B
WebGPU Nodes	548.93 152.19	548.92 152.19	-16 B +0 B

🌳 Bundle size after tree-shaking

Minimal build including a renderer, camera, empty scene, and dependencies.

	Before	After	Diff
WebGL	467.34 112.95	467.34 112.95	+0 B +0 B
WebGPU	624.49 169.06	624.48 169.05	-16 B -14 B
WebGPU Nodes	579.34 158.35	579.33 158.36	-16 B +4 B

[sunag]

Do we have any source that .length=0 is better than other alternatives?

I didn't do an extensive beachmark, but I had done some tests between creating a new array and using .length=0 and creating a new array worked better. It seems that both force the GC, using null values ​​after use was what worked best for small arrays.

[Mugen87]

I've tried to find the original issue/PR where we discussed this but without success so far. It's many years ago but at that time, setting a zero length was the fastest way to clear an array. And it does no create a new array object like the previous approach which invalidates the previous object.

I'm surprised to hear that .length=0 triggers the GC. The array itself is just empty and the reference to it is not set to null or cleared.

[Mugen87]

We can revisit the array management at any point again. Merging for now so the toggling issue is fixed on dev.