What are the trade-offs between precision and recall in model evaluation?