Add model picker: paper's §6.5 token-LM baseline, live-comparable against the thinker

In-browser demo: onnxruntime-web + SentencePiece WASM, 8/8 parity with FINAL_12H

initial commit