fix: unblock event loop for analyze_repo, detach git subprocess stdin on Windows

[huanceng]

• Run analyze_repo via asyncio.to_thread under an asyncio.Lock: keeps tree-sitter usage serialized without blocking the MCP stdio pump (aligns with the to_thread pattern used by all other sync handlers)
• Add stdin=DEVNULL to GitIgnoreFilter's git calls so children never inherit the MCP JSON-RPC stdin pipe, which hangs on Windows ProactorEventLoop

[anhnh2002]

Thanks for the PR, the motivation is right and the stdin change is good hygiene. Moving analyze_repo off the main thread did surface a latent bug though, so I pushed two follow-up commits to your branch.
The per-file parse loop wraps each file in a SIGALRM-based timeout() context manager. signal.signal() can only be called from the main thread, so inside asyncio.to_thread it raised, and the per-file except Exception swallowed that at DEBUG level. Every file was skipped and analyze_repo returned 0 components on macOS and Linux. You would not have seen this on Windows, where SIGALRM does not exist and the helper already falls back to no timeout.

• First commit: timeout() now skips signals when they are unavailable or when not on the main thread, so the analysis runs without a per-file timeout in the MCP server, matching what Windows already did. Also the blank line ruff format wanted and a reworded lock comment, since the lock is really serialising a heavy job rather than protecting tree-sitter across threads.
• Second commit: applied the same stdin=DEVNULL treatment to the remaining git calls in cloning.py and github_processor.py, so no git child inherits the parent's stdin anywhere.