Fine-Tuning Foundation Models with Federated Learning for Privacy Preserving Medical Time Series Forecasting