WebGLBackend: Fix depth regression.

[Mugen87]

Fixed #30526.

Description

#30413 was intended as a performance optimization for the WebGL backend however there are scenarios where no intermediate framebuffer is used (e.g. when using linear-srgb as output color space). Because color space and tone mapping settings can change at any time, the context must honor the initial renderer attributes.

Unfortunately, that means most apps will have a depth buffer attached to the default framebuffer although it isn't required.

[github-actions]

📦 Bundle size

Full ESM build, minified and gzipped.

	Before	After	Diff
WebGL	336.42 78.33	336.42 78.33	+0 B +0 B
WebGPU	518.01 143.8	518.03 143.8	+25 B +3 B
WebGPU Nodes	517.48 143.69	517.5 143.69	+25 B +2 B

🌳 Bundle size after tree-shaking

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

	Before	After	Diff
WebGL	465.41 112.17	465.41 112.17	+0 B +0 B
WebGPU	589.88 159.76	589.91 159.77	+25 B +3 B
WebGPU Nodes	545.25 149.23	545.27 149.23	+25 B +3 B

[Mugen87]

@sunag I wonder if it wouldn't better to always have a framebuffer target, even when using no tone mapping and color space conversion. That would make the setup of the default framebuffer easier and a bit more performant for most use cases (since we could save depth and stencil and MSAA).