On the Trade-off Between Efficiency and Precision of Neural Abstraction