Move quality computation and dataset metrics into indexer

This ensures consistency w.r.t. deduplication and prepares for using global information during quality scoring while avoiding the costs of a second pass in the harvester as we have to read in all datasets for indexing in any case.

Closes #357 (closed)