[codex] Update serving SOTA loop framework docs

[BBuf]

Summary

• Add TokenSpeed and TensorRT-LLM PR optimization history folders with Qwen3.5 and Kimi bilingual diff-backed dossiers.
• Extend llm-serving-auto-benchmark and llm-torch-profiler-analysis for TokenSpeed/TensorRT-LLM flows, official-container runbooks, cookbook configs, and normalized comparison output.
• Update sglang-sota-humanize-loop so SGLang is the fixed target while comparison_frameworks is caller-controlled, including vLLM-only comparison support.
• Refresh the root README and SGLang SOTA SVG to describe selected comparison frameworks instead of a fixed SGLang/vLLM/TRT-LLM set.

Validation

• xmllint --noout docs/assets/sglang-sota-performance-loop.svg
• git diff --check --cached
• python3 skills/llm-serving-auto-benchmark/scripts/validate_cookbook_configs.py skills/llm-serving-auto-benchmark/configs/cookbook-llm
• python3 -m unittest discover tests
• python3 model-pr-optimization-history/scripts/query.py --list

Notes

This keeps compiled-output review out of the SOTA loop, per request, and focuses the loop on selected serving competitors, profiler evidence, and PR-history-backed source choices.