Prevent AuthKeyDuplicatedError from concurrent server instances

[marksmeed]

Summary

• MCP clients (Claude Desktop in particular) can spawn a second telegram-mcp process before the first one exits, e.g. when restarting a connector. Two processes connecting the same Telegram session at once trips AuthKeyDuplicatedError, which can knock out both connections rather than just the newcomer.
• Adds a per-session, cross-platform (flock/msvcrt) advisory lock (telegram_mcp/singleton.py) acquired before connect(). A second instance racing for the same session waits briefly (default 20s, TELEGRAM_LOCK_GRACE_SECONDS) for the first to release it — covering the connector-restart case — and otherwise exits without ever calling connect(), so the live session is never disturbed.
• This sits in front of the existing retry-with-backoff AuthKeyDuplicatedError handling in _connect_authorized_client, which is preserved as-is for genuinely transient blips (e.g. a VPN IP change) that aren't caused by a second local instance.

Test plan

• uv run pytest tests/test_runner.py -v — 11 passed, including new tests for lock contention/release and independent sessions not blocking each other
• uv run pytest — no new failures introduced (pre-existing failures on this Windows checkout are environment-/platform-specific and unrelated: multi-account fixtures assuming a single account, and tests/test_session_pool.py importing POSIX-only fcntl)