fix: guard litellm modify_params context

[he-yufeng]

Summary

• serialize litellm.modify_params save/set/restore with an RLock
• cover the context manager with a threaded regression test
• keep existing retry/completion paths unchanged

Fixes #3137.

To verify

• python -m py_compile openhands-sdk\openhands\sdk\llm\llm.py tests\sdk\llm\test_llm_completion.py
• python -m ruff check openhands-sdk\openhands\sdk\llm\llm.py tests\sdk\llm\test_llm_completion.py
• python -m ruff format --check openhands-sdk\openhands\sdk\llm\llm.py tests\sdk\llm\test_llm_completion.py
• python -m pytest tests\sdk\llm\test_llm_completion.py -q

[he-yufeng]

Rebased this branch onto current main and resolved the LLM token-limit conflict by keeping the new _effective_max_* private attrs alongside the modify_params lock.

Revalidated locally after the rebase:

• .venv\Scripts\python.exe -m pytest tests\sdk\llm\test_llm_completion.py -q --basetemp C:\dev\GITHUB-clean\_tmp\software-agent-sdk-pytest - 17 passed
• .venv\Scripts\python.exe -m ruff check openhands-sdk\openhands\sdk\llm\llm.py tests\sdk\llm\test_llm_completion.py
• .venv\Scripts\python.exe -m ruff format --check openhands-sdk\openhands\sdk\llm\llm.py tests\sdk\llm\test_llm_completion.py
• .venv\Scripts\python.exe -m py_compile openhands-sdk\openhands\sdk\llm\llm.py tests\sdk\llm\test_llm_completion.py
• git diff --check

[enyst]

Thank you for the contribution! Calvin's investigation flagged this and wrote the tracking issue. @csmith49 I would love it if you could maybe take a look at the PR?

[all-hands-bot]

🟢 Good taste - Clean, minimal fix for a real concurrency bug.

[RISK ASSESSMENT]
🟢 LOW - Internal synchronization fix with no API or behavior changes. The RLock correctly serializes access to global litellm state, and the regression test thoroughly validates the thread-safety guarantee.

VERDICT:
✅ Worth merging - Straightforward concurrency fix with proper testing.

KEY INSIGHT:
Using a ClassVar RLock is exactly right here since litellm.modify_params is global state that needs process-wide synchronization.