Feature Distillation is the Better Choice for Model-Heterogeneous Federated Learning