fix(mcp): support stdio transport on Windows

[Sanjays2402]

Description

• Fixes [Bug] MCP server stdio transport crashes silently on Windows (add_signal_handler raises NotImplementedError) #7595. Windows asyncio loops raise NotImplementedError when registering signal handlers, so stdio now treats that registration as optional and continues serving the MCP client.
• No dependencies required.

How has this been tested?

• python -m pytest tests -q — 174 passed, 4 skipped.
• Regression test simulates an event loop without signal-handler support and verifies the stdio server still starts.
• ruff check passes on both changed files.

Checklist

• I have performed a self-review of my own code.
• I have commented the Windows-specific behavior.
• I have adhered to the GitFlow naming convention (hotfix/mcp-stdio-windows).
• I am following the contributing guidelines.
• Tests cover the bug fix.

[CLAassistant]

All committers have signed the CLA.

[DogInfantry]

Tested this on Windows 11 (build 26200). Confirmed the root cause: on a ProactorEventLoop, loop.add_signal_handler(signal.SIGINT, ...) raises NotImplementedError with an empty str(), so stdio dies before the handshake. With this branch's try/except NotImplementedError the server proceeds and serves over stdio. The full mcp_server test suite passes here (176 passed, 2 skipped), including the new test_stdio_main_runs_when_signal_handlers_are_unsupported. LGTM on the crash fix.

One note: this covers the crash, but not the second problem #7595 raises. main()'s handler logs logger.error("Server error: %s", e), and since str(NotImplementedError()) is empty the user sees a blank Server error: with no type and no traceback. I opened a small complementary PR (#7600) that switches it to logger.exception and includes the exception type, so future fatal errors are diagnosable. Happy to fold it into this PR instead if you would prefer a single change.

[deeleeramone]

Thanks for the fix!