[WebGPU port] Backend-neutral vertex layout descriptor

[obiot]

Tracking ticket — defer to actual WebGPU port

A small refactor identified during 19.7 prep that's deliberately not being landed in 19.7 because the right shape depends on a real WebGPU consumer.

Today (imperative WebGL)

Each batcher (QuadBatcher, LitQuadBatcher, MeshBatcher, PrimitiveBatcher, and the tilemap GPU renderer) configures its vertex layout by emitting a sequence of gl.vertexAttribPointer calls — see Batcher.setVertexAttributes(gl, attributes, stride):

gl.vertexAttribPointer(0, 3, gl.FLOAT, false, stride, 0);   // aVertex (vec3)
gl.enableVertexAttribArray(0);
gl.vertexAttribPointer(1, 2, gl.FLOAT, false, stride, 12);  // aRegion (vec2)
gl.enableVertexAttribArray(1);
gl.vertexAttribPointer(2, 4, gl.UNSIGNED_BYTE, true, stride, 20); // aColor (uint8x4 normalized)
gl.enableVertexAttribArray(2);

WebGPU equivalent (declarative)

const layout = {
  arrayStride: 24,
  attributes: [
    { shaderLocation: 0, format: "float32x3", offset: 0  },  // aVertex
    { shaderLocation: 1, format: "float32x2", offset: 12 },  // aRegion
    { shaderLocation: 2, format: "unorm8x4",  offset: 20 },  // aColor
  ],
};

Same information, different shape. WebGPU's descriptor is passed at pipeline-creation time, not per-draw.

Refactor sketch

A backend-neutral descriptor each batcher carries once:

const QUAD_VERTEX_LAYOUT = {
  stride: 24,
  attributes: [
    { name: "aVertex", type: "f32x3", offset: 0  },
    { name: "aRegion", type: "f32x2", offset: 12 },
    { name: "aColor",  type: "u8x4n", offset: 20 },  // n = normalized
  ],
};

• WebGL backend: helper iterates the descriptor and emits the matching gl.vertexAttribPointer calls.
• WebGPU backend: helper translates to the declarative arrayStride / attributes shape above.

Why deferred

The right type vocabulary (f32x3 / unorm8x4 / etc.) depends on:

1. What WebGPU's GPUVertexFormat enum actually supports.
2. What WebGL formats each batcher uses today (need to audit Batcher.setVertexAttributes calls and the supplied attributes shape).
3. What the future WebGPU equivalent of MaterialBatcher would need.

Designing the descriptor vocabulary up-front, without a real WebGPU consumer to validate against, risks committing to a shape that needs reshuffling. Better to land this as one of the first commits of the actual WebGPU port — that way the abstraction gets validated against a real implementation rather than designed in the abstract.

Cross-refs

• Refactor TextureCache and Batcher base class for renderer-agnostic abstraction #1410 (TextureCache + Batcher base class refactor — 20.x candidate, this would slot in naturally there).
• Each batcher source file under src/video/webgl/batchers/ carries its own setVertexAttributes call shape — audit those first when picking the descriptor type vocabulary.