Scaling Law Analysis in Federated Learning: How to Select the Optimal Model Size?