Fixes framebuffer caching with multiple references of different sizes

[arodic]

#Change summary

The Root Cause of #32689

• ViewportDepthTextureNode: Uses a singleton _sharedDepthbuffer that is shared across all RenderTargets/CanvasTargets. This causes brute-force destruction and re-creation of depth buffer per reference on every rendered frame if references are not exactly the same size. This affects both CanvasTarget and RenderTarget

• ViewporTextureNode Does not consider CanvasTarget as a reference. It simply does not do caching for CanvasTargets

• Renderer: Uses a single _frameBufferTarget that resizes for each canvas (if not exactly the same size), triggering
  depthTexture.needsUpdate = true and causing texture recreation

The Fix for #32689

• ViewportDepthTextureNode removed getTextureForReference() override to allow texture caching. Now it caches depth textures per reference

• ViewportTextureNode Added check for CanvasTarget. Previous implementation considered RenderTargets only.

• Renderer Changed _frameBufferTarget to _frameBufferTargets (Map<CanvasTarget, RenderTarget>). Each canvas target now gets its own framebuffer target instead of sharing one

This allows multi-canvas rendering with different sizes and tone mapping enabled without constant texture buffer recreation.

[github-actions]

📦 Bundle size

Full ESM build, minified and gzipped.

	Before	After	Diff
WebGL	357.49 84.84	357.49 84.84	+0 B +0 B
WebGPU	621.89 172.77	622.07 172.79	+176 B +21 B
WebGPU Nodes	620.49 172.53	620.67 172.55	+176 B +18 B

🌳 Bundle size after tree-shaking

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

	Before	After	Diff
WebGL	490.27 119.86	490.27 119.86	+0 B +0 B
WebGPU	692.79 188.01	692.97 188.03	+176 B +21 B
WebGPU Nodes	642.17 175.11	642.35 175.13	+176 B +23 B

[arodic]

I cleaned up all the slop. This should be minimal and pretty clean implementation fixing the bug. Ready for review!

[arodic]

I also added tests (generated by Opus 4.5) ViewportDepthTextureNode.tests.js‎

Let me know if this is needed. From what I can tell, it tests for the correct edge case.

[sunag]

Could we add an individual dispose event for each CanvasTarget?

[Mugen87]

While working on some viewport/scissor related logic, I have found out this PR broke webgpu_multiple_canvas. You see the same mesh everywhere instead of different ones like in prod:

https://threejs.org/examples/webgpu_multiple_canvas

Right before this PR, dev was fine.

[arodic]

@Mugen87 interesting. Ili take a look. Curious why screenshot tests didn’t pick this up?

[Mugen87]

webgpu_multiple_canvas is on the exception list because it's a WebGPU only demo and the CI relies on WebGL 2 (it does not support WebGPU yet).

[arodic]

@Mugen87 I drafted two PRs #33209 and #33210

Personally I'm leaning towards the latter. Lmk wdyt?

[Mugen87]

I would prefer if @sunag decides for a solution here 😇 . He knows this part of the library best.

[sunag]

@arodic Thanks for the PRs, I implemented a solution similar to yours but caching the Quads; the problem with caching the nodes after all the changes was the need to revalidate the shader, which made it a bit less efficient.