Added wiki-gemma dataset and updated data distribution to only limit swe/no datasets to 10k

Added train-test split, and reduced each dataset to max 10k to balance the dataset sizes (DK datasets was underrepresented)