Harvester metrics are inconsistent with deduplicated datasets

Harvester metrics like datasets per license or per tag are computed on the fly before deduplication. Since deduplication will remove some of the dataset, the metrics can be inflated, for example today the tag "Behörde" has 70 datasets after deduplication but the metrics not 101 dataset which make the tag eligible for the top-N computations. So we either need to fix up the detailed metrics like we already do for the overall or we need to collect the metrics in a separate pass over all datasets.