pub fn benchmark_devices(n: usize) -> CapabilityMatrixExpand description
Benchmark every available compute circuit and return the ranked capability matrix.
n is the GEMV side length (representative shape; 2048 is a good fast default).
Each (vendor, device, backend) triple is benchmarked separately so DX12 vs Vulkan
(same physical GPU) can rank against each other — the whole point of the passport.
The software/WARP “CPU” wgpu adapter is skipped — the native rayon path is the honest CPU number.