Improve finalization of the disk-based aggregation generations

Currently, during the finalization process, in order to merge data from one generation to another, the hash map of one generation is iterated. After that, the corresponding row in the other generation is searched and the aggregated data are merged. This algorithm is quite complex and inefficient in terms of memory usage and disk operations. It can be simplified by iterating not by the hash map, but by the rows in the vector of RGDatas. This would require slightly more memory to duplicate the row hashes in an RowStorage (+8 bytes per row). However, it should speed up the process and reduce peak memory consumption.