PUMA: Efficient Continual Graph Learning with Graph Condensation