Modeling issue 89 modeling template cycle 3 random forest

[lima-tango]

Done:

• Use only 4 negative labels and two additional (needsmoderation and negative) labels.
• Balance data with translation and oversampling, both with sampling strategy 0.9 only.
• Use only normalization
• Update cycle-tag
• Regularize GridSearch params: max_depth and min_samples_leaf -> Overfitting reduces the loss probably due to oversampling and translation in cross-validation split

Closes #89