Taming Cross-Domain Representation Variance in Federated Prototype Learning with Heterogeneous Data Domains