fix: handle CancelledError in MCP tool calls to prevent process crash

[pixan-ai]

What

Add proper exception handling for CancelledError and general Exception
in MCPToolWrapper.execute().

Why

MCP SDK uses anyio cancel scopes internally. When a tool call times out or
fails, anyio can produce a CancelledError that leaks through asyncio.wait_for.

Since CancelledError inherits from BaseException (not Exception in Python 3.9+),
it escapes both MCPToolWrapper.execute() (which only caught TimeoutError) and
ToolRegistry.execute() (which only catches Exception), propagating up to the
agent loop dispatch and crashing the process.

The existing close_mcp() method already acknowledges this MCP SDK behavior
by catching BaseExceptionGroup.

How

• Catch asyncio.CancelledError: use task.cancelling() (Python 3.11+) to
  distinguish genuine cancellation (/stop) from MCP SDK internal cancellation.
  Re-raise for /stop, return graceful error otherwise.
• Catch Exception: return error message to LLM instead of crashing.

Testing

1. Configure an MCP server with a slow tool
2. Set a short toolTimeout
3. Trigger the tool → should return graceful error instead of crashing
4. Use /stop during MCP tool execution → should still cancel properly

Related: #1055

Acknowledgment

Fix identified and developed with assistance from Claude Opus 4.6