The Risk of Federated Learning to Skew Fine-Tuning Features and Underperform Out-of-Distribution Robustness