Latte: Collaborative Test-Time Adaptation of Vision-Language Models in Federated Learning