test: make Gemini async integration test robust across pytest async modes

[shuofengzhang]

What changed

• Updated tests/clients/test_gemini.py integration test test_async_completion from an async def pytest test to a synchronous test that calls asyncio.run(...).
• Added import asyncio for the wrapper call.

Why

• The integration test is guarded by GEMINI_API_KEY, so it is often skipped in CI/local runs.
• When the key is present and the test actually runs, pytest can fail with:
  • async def functions are not natively supported
• This happens because the test relies on async test plugin behavior that can vary by pytest/pytest-asyncio mode.

Insight / Why this matters

• Root cause: the test used plugin-managed async execution for a single integration case that can be skipped most of the time, so breakage is easy to miss until someone runs with real credentials.
• Why easy to miss: default/CI environments without GEMINI_API_KEY skip this path, masking the failure mode.
• Tradeoff: asyncio.run(...) makes the test self-contained and independent of async test plugin mode; this slightly reduces flexibility for nested event-loop contexts, but for this integration test in a normal pytest process it is the safest default.
• Impact: contributors with GEMINI_API_KEY can run the suite without hitting a framework-level async test failure unrelated to client correctness.

Practical gain / Why this matters

• Improves test reliability and developer UX for real-key integration runs.
• Removes an avoidable source of false-negative failures in local verification.
• Keeps scope rollback-safe (single test change, no runtime package logic changed).

Testing

• Reproduced failure before change:
  • scripts/clone_and_test.sh alexzhang13/rlm
  • Result included: FAILED tests/clients/test_gemini.py::TestGeminiClientIntegration::test_async_completion
• Verified after change:
  • scripts/clone_and_test.sh alexzhang13/rlm
  • Result: 277 passed, 7 skipped