fix(mcp): reuse stdio watcher coroutine

[Serph91P]

Summary

• create the stdio child-watcher coroutine exactly once
• reuse that coroutine for awaitability validation and task scheduling
• prevent the discarded probe coroutine from emitting RuntimeWarning: coroutine ... was never awaited
• assert in the reconnect regression test that the watcher is constructed once

The MCP-prefixed pre-discovery validation changes were removed because they are already covered more broadly by #44044.

Verification

• uv run pytest -q tests/tools/test_mcp_stdio_fastfail_reconnect.py (2 passed)
• uv run pytest -q tests/tools/test_mcp*.py (596 passed)
• uv run ruff check tools/mcp_tool.py tests/tools/test_mcp_stdio_fastfail_reconnect.py
• git diff --check

[alt-glitch]

This was generated by AI during triage.

Duplicate of #44044 for the configured mcp-<server> pre-discovery validation repair; #44044 is the broader open implementation across startup validators. This PR also carries separate stdio watcher-warning cleanup, which should be preserved if work is consolidated.

[Serph91P]

Thanks for flagging this. I removed the duplicate MCP-prefixed validation changes. This PR now contains only the independent stdio watcher coroutine cleanup; #44044 remains the implementation for pre-discovery validation.

[teknium1]

Closing with credit — the leak this fixes landed on main via #100920 (fe3e5dc), carrying @loulanyue's earlier #96044 fix (Aug 27) with a different shape: the probe now uses inspect.iscoroutinefunction(_watch_children) instead of creating the coroutine once and reusing it. Same bug, same outcome (0 leaked coroutines, verified live). Thanks @Serph91P for the independent fix.