docs: record the CPU TinyLlama HumanEval pair

[Mattral]

Summary

Records the CPU float32 TinyLlama pair measured in this session at git f59b5b4.

• Target TinyLlama/TinyLlama-1.1B-Chat-v1.0, drafter Felladrin/Llama-160M-Chat-v1, HumanEval/23, 16 new tokens, n_warmup=0, n_samples=1, gamma 4, one thread.
• Command: python benchmarks/run_all.py --config <yaml outside benchmarks/configs/> --output-dir <dir>, with OMP, MKL, OPENBLAS, and NUMEXPR threads set to 1.
• FlashSpec 1.7507264092133707 tok/s, vanilla autoregressive 4.2977579080394 tok/s, speedup_vs_ar_mean 0.40735807988124606, acceptance 0.11363636363636363.
• The README opening says the engine is currently slower than plain decoding, and the tagline no longer says the verification kernel is Triton-optimised.
• The Colab Tesla T4 JSON files were already in the tree (4e48c55). The status box now cites benchmarks/results/tinyllama_gpu_flashspec_ucb_cb02ca1.json and benchmarks/results/tinyllama_gpu_vanilla_ar_cb02ca1.json. That run is a recorded measurement.
• The H100 tok/s table removed earlier was an unmeasured design-target table. Those figures were never in a result JSON. paper/flashspec.tex still labels the same table as unmeasured. This PR does not restore those numbers and does not call that table a deleted measurement.

flashspec/sampling/rejection.py is unchanged, so merging should not queue asm-residual.yml.