Federated Full-Parameter Tuning of Billion-Sized Language Models with Communication Cost under 18 Kilobytes