Cuing Without Sharing: A Federated Cued Speech Recognition Framework via Mutual Knowledge Distillation