fix(sglang): surface dropped rollout arguments

[CjhHa1]

Summary

• warn when HTTP/native SGLang backends drop unknown ServerArgs keys, with UNIRL_SGLANG_STRICT_SERVER_ARGS=1 available to fail closed
• document the Qwen3 colocate full-FT, async/separate, LoRA, memory, CUDA graph, weight-sync, and reserved-port contracts
• audit the in-tree Qwen3 SGLang recipes and annotate their preset or intentional deviation without changing effective values

Related Issue

Fixes #261

Test Plan

• SKIP=no-commit-to-branch uvx pre-commit run --all-files --show-diff-on-failure — passed
• CPU assertion harness for warn/strict dropped-key behavior plus server_intent() precedence and reserved-port overrides — passed (20 focused assertions during development, followed by a compact reproducible smoke)
• GPU rollout smoke: Not run; this changes boot validation and documentation without changing effective recipe values. The repository policy removes committed unit-test harnesses, so the CPU checks remain one-off verification.

Compatibility / Risk

Default behavior remains permissive: unknown SGLang keys are filtered as before but now emit a warning. Strict failure is opt-in through UNIRL_SGLANG_STRICT_SERVER_ARGS=1. UniRL-only transport keys are excluded from warnings. No checkpoint, data-format, recipe-value, or resource-requirement changes.

Reviewer Notes

AI-assisted implementation; the submitter reviewed the full diff. Duplicate-work check found no open PR addressing #261 or SGLang ServerArgs key-drop validation.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.