Vanilla relative filters

[roeysc]

Consider adding another filter in the FeatureExtractor.py, for excluding games with too many gallery shapes.
Currently Both Yogev and Roey are using such a filter downstream (e.g., before running Pareto), but that isn't the right way to go, because all the relative features are calculated relative to the vanilla dataset that does include these bad games.
This change would require adding a mask in FeatureExtractor._get_sample_relative_filters, as well as relevant constants in behavioral._consts. The default_config.yml can stay the same assuming that we use the same threshold of 3 Z-scores.

[roygutg]

I think there's an absolute filter that might do a similar job - #6.
If it's implemented and used to filter games, could that solve this issue?

[roeysc]

It will probably give very similar results, yes, assuming that an extreme number of gallery shapes must include many step-save pairs. It's still possible to get too many gallery shapes without such step-save pairs, so I think it's still worth adding.