Communication-Efficient Personalized Federal Graph Learning via Low-Rank Decomposition