High-fidelity digital mask optimization using a potential-driven multi-learning particle swarm optimizer with consecutive-descent guidance