Pool the WebGPU readback staging buffers

WebgpuGraphicsDevice.readStorageBuffer creates a staging buffer for every read and readBuffer destroys it once the read settles, so a readback issued every frame — which SceneDepthReader does, and Picker can — creates and destroys a GPU buffer every frame. There is no pooling anywhere on this path.

https://github.com/playcanvas/engine/blob/main/src/platform/graphics/webgpu/webgpu-graphics-device.js#L1497

Pooling these is safe and needs no API change, because the buffer never escapes: readBuffer does getMappedRange → data.set(...) → unmap(), copying the bytes into the caller's array, so the buffer's lifetime is entirely internal. That is unlike the arrays handed back to callers, which deliberately are not pooled — the device cannot know when a caller has finished reading one.

Notes for whoever picks this up:

• A mapped buffer cannot be reused until it is unmapped, so the pool has to hand back only unmapped buffers, and key them by size.
• webgpu-upload-stream.js already manages staging buffers for uploads and is worth mirroring, or extending.
• core/object-pool.js is not suitable: it is a bulk-reset pool (allocate() plus freeAll(), with no individual release) and constructs with no arguments, so it cannot size a buffer.

Came out of #9293, which fixed a WebGL-only readback stall by moving the blocking copy to the start of the next frame. WebGPU was never affected by that, as mapAsync does not block — this is a separate, backend-specific waste.