NodeMaterialObeserver: Fix detection of replaced attribute/index.

[Mugen87]

Fixed #32926.

Description

Ensures NodeMaterialObeserver can detect replaced geometry attributes and indices.

[github-actions]

📦 Bundle size

Full ESM build, minified and gzipped.

	Before	After	Diff
WebGL	359.21 85.29	359.21 85.29	+0 B +0 B
WebGPU	618.3 172.12	618.4 172.15	+98 B +24 B
WebGPU Nodes	616.88 171.87	616.98 171.9	+98 B +23 B

🌳 Bundle size after tree-shaking

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

	Before	After	Diff
WebGL	490.85 119.64	490.85 119.64	+0 B +0 B
WebGPU	692.12 187.28	692.22 187.3	+98 B +22 B
WebGPU Nodes	641.5 174.69	641.6 174.71	+98 B +22 B

[Mugen87]

@sunag Unrelated to this PR but since the other PR modified below code section:

three.js/src/materials/nodes/manager/NodeMaterialObserver.js

Lines 590 to 596 in ca544b9

	if ( this.renderId !== renderId ) {
	this.renderId = renderId;
	return true;
	}

This bit is actually a bandage and should be removed from NodeMaterialObserver. It is currently required though to make sure "frame" and "render" scoped bindings are at least updated once. This is a design issue that should be resolved.

For that, I believe we have to build the node materials after the render lists are ready but before _renderObjectDirect() is called. Then the "frame" and "render" bindings (e.g. the camera related uniforms) are updated just once. _renderObjectDirect() should then only update the object related bindings.

The node material build is currently triggered in needsRefresh() which is too late.

Besides, "frame" and "render" bindings are unnecessarily updated too often. UniformsGroup can detect that no update is required beause of the internal value cache but the checks itself are not necessary in the first place.

To sum up, the renderer should do this:

• Prepare render lists.
• Build all node materials in the lists by calling NodeManager.getForRender() for all render objects.
• Update frame and render scoped bindings once if necessary.
• Iterate through render lists, call _renderObjectDirect(), check for refresh and update object scoped bindings if necessary.

[arodic]

This seems like a pretty proper fix to me. Thanks! Any feedback on #32689/#32690?

[Mugen87]

I'm trying to have a look at this when the bind group cache is fixed. This is a performance issue that affects all WebGPURenderer applications right now (see #32913 and #32934).

Besides, I'm not sure what design plans @sunag has with the viewport nodes and CanvasTarget. Maybe he wants to solve #32689 differently so I would prefer to wait on his feedback.