Try instance-rate vertex buffer for per-instance data for CPU-driven rendering

## Objective

Improve performance of 2D/3D mesh rendering.

## Solutions

I observed a significant performance boost to sprite rendering when moving to using an instance-rate vertex buffer for per-instance data, compared to using a uniform or storage buffer, such as via `GpuArrayBuffer`.

Because of the constraints on CPU-driven rendering and not being able to feasibly implement meshlet-based rendering on the CPU to be able to batch draws across different meshes, there is no need to be able to randomly access per-instance data, and we anyway write it in draw order.

As such, it seems reasonable to try out using an instance-rate vertex buffer for per-instance data to see if it brings a performance benefit to 2D/3D mesh rendering.

Start by tracing CPU-driven rendering and the existing per-instance data path, including uses of GpuArrayBuffer, then compare it with the sprite-rendering path described in the issue. The work is done when instance-rate vertex-buffer data is implemented for the relevant 2D/3D mesh rendering path and measurements show whether it improves performance.

Written by the indexing model from the issue text.