fix(quality): documented llama-server backend and run_qwen36_benchmark.sh are missing

[kurosawareiji7007-hub]

Summary

bench/quality/README.md documents the quality-parity A/B path as:

python3 run_quality.py --backend llama-server --llama-url http://localhost:8082 ...

and references orchestrator bench/quality/run_qwen36_benchmark.sh.

Neither exists today:

• run_quality.py only accepts oracle|mock|sparkinfer|llama — argparse rejects llama-server
• run_qwen36_benchmark.sh is missing

So the documented workflow for proving “same GSM8K/IFEval/… as llama.cpp at higher tok/s” cannot run.

Expected

1. --backend llama-server + --llama-url against llama.cpp /completion (fallback /v1/completions)
2. run_qwen36_benchmark.sh runs sparkinfer then llama-server exclusively
3. CPU-only --self-test-llama-server mock to verify wiring without a GPU

Scope

bench/quality/ only (not bench/scripts/). No GPU required to review/verify the self-test.