An improved GNN integrating multi-scale topology-semantic attention for performance prediction in spindle assembly