ROAR dataset is loaded from cache when result is missing

[AndreasMadsen]

For convience all results are tied to the results/*.json files. If you delete one .json file the results can be recomputed. However, right now the ROAR datasets are not recomputed but instead loaded from cache. This is problematic because the ROAR dataset depends on the model results.

This is a minimal problem, in the sense that the encoded-roar dir can just be deleted. However, the roar.py should always rebuild the dataset and save, and only load from cache for k-1.