Model Pruning Enables Localized and Efficient Federated Learning for Yield Forecasting and Data Sharing