Towards Robust and Efficient Federated Low-Rank Adaptation with Heterogeneous Clients