Fix: close MCP sessions after canvas execution to prevent connection leaks

Add Graph.close() method that finds and closes all MCPToolCallSessions
held by Agent components. Call it in finally blocks after canvas.run()
in canvas_service.py and canvas_app.py. Also move session cleanup to
finally blocks in test_tool endpoint and get_mcp_tools to ensure
sessions are closed even on exceptions.

api/apps/canvas_app.py

@@ -226,6 +226,8 @@ async def run():
             logging.exception(e)
             canvas.cancel_task()
             yield "data:" + json.dumps({"code": 500, "message": str(e), "data": False}, ensure_ascii=False) + "\n\n"
+        finally:
+            canvas.close()
 
     resp = Response(sse(), mimetype="text/event-stream")
     resp.headers.add_header("Cache-control", "no-cache")

api/apps/mcp_server_app.py

@@ -367,11 +367,12 @@ async def test_tool() -> Response:
         tool_call_sessions.append(tool_call_session)
         result = await thread_pool_exec(tool_call_session.tool_call, tool_name, arguments, timeout)
 
-        # PERF: blocking call to close sessions — consider moving to background thread or task queue
-        await thread_pool_exec(close_multiple_mcp_toolcall_sessions, tool_call_sessions)
         return get_json_result(data=result)
     except Exception as e:
         return server_error_response(e)
+    finally:
+        # PERF: blocking call to close sessions — consider moving to background thread or task queue
+        await thread_pool_exec(close_multiple_mcp_toolcall_sessions, tool_call_sessions)
 
 
 @manager.route("/cache_tools", methods=["POST"])  # noqa: F821