[FIX] Several fixes in learners/models

[PrimozGodec]

Issue

• Logistic regression uses MajorityLearner when only one class in data, while SVM fails
• Softmax regression and NuSVM return the wrong shape of the probability matrix when one class missing in training data.
• Now when the data_to_model_domain was introduced in a Model class it is duplicated in outliers.
• Rules fail when empty list with new rules

Description of changes

• LogisticRegressionLearner does not use MajorityLearner when only one class in data.
• Softmax regression fixed to return the right shape of probability matrix (num cases x num classes in the domain)
• NuSVM: removed own predict function, adding missing dimensions is now handled by predict function on SklModel
• data_to_model_domain removed from outliers, added callbacks to data_to_model_domain in Model class
• Rules fixed

Includes

• Code changes
• Tests

[codecov]

Codecov Report

Merging #4655 into master will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff            @@
##           master    #4655    +/-   ##
========================================
  Coverage   83.69%   83.69%            
========================================
  Files         280      275     -5     
  Lines       56447    55604   -843     
========================================
- Hits        47244    46539   -705     
+ Misses       9203     9065   -138