[Hotfix] CV unbalanced subsampling

[itellaetxe]

Fixes #76

When training a logistic regression classifier for 2 unbalanced classes using age deltas or features (or whatever), we use majority class subsampling. We used to do it before going into the Cross-Validation procedure, leading to potential optimistic biases and underutilization of the data.

Now we put all the data available in the CV, and we undersample the majority class for each CV fold, avoiding the described problem.

[JGarciaCondado]

Please change to always subsampling, not based on a ratio. Open an Issue saying that the user should be allowed to change this if they want. Explain why the specific seed was set.