Benchmarking on Tasks That Matter: Dataset Selection for Preserving Model Rankings