fix: lazily reconnect MCP tools on call

[WittmannF]

Summary

• Adds a best-effort lazy reconnect path when an MCP tool schema exists but the in-process server/session is missing
• Removes stale disconnected server entries before rediscovery so configured servers can reconnect without a manual gateway restart
• Adds regression coverage for missing-server lazy reconnect on tool calls

Test Plan

• python -m pytest tests/tools/test_mcp_tool_session_expired.py -o 'addopts=' -q
• python -m pytest tests/tools/test_mcp_tool.py tests/tools/test_mcp_tool_session_expired.py tests/tools/test_mcp_oauth*.py tests/hermes_cli/test_mcp_config.py -o 'addopts=' -q

[teknium1]

Closing with credit — you were the earliest (May 11) to identify this bug class: a dead cached MCP server entry that a new session can't recover. The fix landed differently: the run task now never dies (parks + self-probes every 5 min, PR #59222), tools re-register on revival, and register_mcp_servers nudges any cached entry with a dead transport at session startup (PR #59331, merged). Your handler-side lazy reconnect would now race the always-alive task, so it wasn't taken directly, but the problem you flagged is fixed. Thanks!