Communication-Efficient and Tensorized Federated Fine-Tuning of Large Language Models