Towards Understanding and Mitigating Dimensional Collapse in Heterogeneous Federated Learning