SAGE: Scalable Ground Truth Evaluations for Large Sparse Autoencoders