docs(qwen35): add 9B and 27B single-GH200 serving numbers

[FeathBow]

Description

• Adds Qwen3.5-9B Serving Load and Qwen3.5-27B Serving Load to /models/qwen35-4b/, measured on 1x GH200 120GB (aarch64, sm_90) during the openinfer 509/510 rerun campaign: full c/QPS ladders with TTFT/TPOT p50+p99, long-context and c=120 overload behavior, tool-call round-trips, and greedy-vs-HF plus per-size logits-gate summaries.
  • 9B: openinfer main ffb959c4, Python vllm bench serve client, random 1024/128, greedy, default seed. Saturation ~6.7 req/s / ~857 out tok/s.
  • 27B: ffb959c4. Plateau ~2.7 req/s / ~340 out tok/s at qps≥8; the load-time budget line is quoted and explained — 17.2 GB KV, 5.5 GB prefill scratch, 18.8 GB recurrent-state reserve (two ~147 MB states per decode slot across the 64-slot capacity).
• Rewrites the Notes section for the family: per-request recurrent-state sizes (~49 MB at 9B / ~147 MB at 27B), the load-time reserve and the --max-batch capacity/KV trade-off, the decodable-vocab selection bound (248077 of the 248320 padded lm_head rows), and a corrected CUDA-Graph statement — the server always captures and rejects --cuda-graph=false, so the previous "disable for debugging" note no longer holds.
• Links the overview table row: Qwen3.5-4B / 9B / 27B → /models/qwen35-4b/.
• Validation: npm run build + static tests 12/12.

[xiaguan]

LGTM. The Qwen3.5 family coverage, serving-table arithmetic, memory-budget calculations, runtime notes, and existing route coverage are consistent. The full website test suite passes.