fix(grpc): support signal shutdown on Windows

[primorLee]

Why are these changes needed?

GrpcWorkerAgentRuntime.stop_when_signal() currently calls loop.add_signal_handler() unconditionally. The default Windows event loop does not implement that API, so cross-language workers fail with NotImplementedError instead of waiting for shutdown. The gRPC host has the same signal-registration path.

This change keeps native asyncio signal handlers on supported event loops and falls back to signal.signal() when the event-loop API is unavailable. The fallback schedules shutdown on the owning loop with call_soon_threadsafe, and both paths remove or restore their handlers when waiting finishes. A shared helper keeps worker and host behavior consistent.

Related issue number

Closes #5760

Checks

• I've included any doc changes needed for https://microsoft.github.io/autogen/. No public API or documentation changes are required.
• I've added tests corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed. CI is pending.

Local validation:

• pytest -q python/packages/autogen-ext/tests/test_grpc_signal_utils.py — 2 passed on Windows
• Ruff 0.4.8 format and lint — passed for all four changed files
• Pyright 1.1.389 — 0 errors for all four changed files
• strict mypy — 0 errors for the new signal helper

AI assistance: Codex was used to inspect the failing paths, implement the fix, and run the checks above. The final diff and test behavior were reviewed before submission.

[Dinesh Yadav (dineshyadav03)]

I read through _signal_utils.py and both call sites. The approach is sound: try add_signal_handler first, catch the NotImplementedError it actually raises on Windows' default event loop, then fall back to signal.signal() with call_soon_threadsafe to safely hop back onto the loop from the signal handler thread. Deduping signals via dict.fromkeys and restoring the previous handlers in finally are nice touches.

One concrete gap: neither new test in test_grpc_signal_utils.py is marked @pytest.mark.windows. This package's CI Windows job runs pytest -m 'windows', and that marker is the established pattern for Windows-only coverage. Without it, these tests execute on the Linux job via a monkeypatched NotImplementedError but are never collected by the actual windows-latest runner — so the fallback path stays CI-unverified on the platform it targets, despite the "2 passed on Windows" note being local-only.

Separately: is SIGBREAK (Ctrl+Break) worth including in the default signal set for stop_when_signal, or intentionally out of scope?

I'm on Windows and can sanity-check the fallback locally if useful.

Requesting changes — logic is correct, but add the windows marker so CI actually exercises the code this PR fixes.

[primorLee]

Addressed the requested Windows CI coverage in 2291f826:

• added @pytest.mark.windows to both new signal-handler tests;
• verified the targeted file on Windows: 2 passed;
• verified the actual Windows selector with pytest -m windows: 2 passed;
• Ruff 0.4.8 check and format checks pass.

SIGBREAK is intentionally out of scope for this PR. Issue #5760 is specifically about add_signal_handler() being unavailable on Windows, so this change preserves the existing default (SIGTERM, SIGINT) behavior. The API still accepts a caller-supplied signal sequence, so Windows callers can explicitly pass SIGBREAK when needed. I can follow up separately if maintainers prefer adding it to the default set.