Not all Minorities are Equal: Empty-Class-Aware Distillation for Heterogeneous Federated Learning