EG2 WGSL engine bench

Embeds 12 test queries with the 4-layer EmbeddingGemma 2 cut three ways: the hand-written WGSL engine, onnxruntime-web on WebGPU, and onnxruntime-web on wasm. Each vector is checked against onnxruntime on CPU. Close other GPU-heavy tabs, press Run, then Copy JSON. The full 24-layer model runs the same comparison; it downloads 174 MB of weights from Hugging Face first.