Exploring Cross-Client Memorization of Training Data in Large Language Models for Federated Learning