Monotonic calibration

[rvashurin]

• Implements four types of normalizers as modules in src/lm_polygraph/normalizers folder
• Rewrites polygraph_normalize script to use new normalizer classes
• Provides self-contained notebook with experiments that compare MSE of confidences obtained with different normalizers when compared to AlignScore of test sets

Each normalizer class implements fit and transform methods, as well as dumps and loads functions for normalizer persistence.

IMPORTANT

This PR breaks the approach for displaying confidence in the demo app. This will need to be remedied in the follow-up PR using new normalizer API.

[rvashurin]

Some comments in manager.py could be cleaned up.

Sure