You have a model registered in the Model Registry in Vertex AI and need to evaluate its performance over multiple versions for A/B testing in production. Which approach will best allow you to manage different versions effectively while ensuring they can be accessed and deployed as needed?