Beyond Top Activations: Efficient and Reliable Crowdsourced Evaluation of Automated Interpretability