Release 0.2.0

Refactored severity_model.py to use a manual inference in order to have full control over performance, consistency, and debugging. pytorch is directly used instead of the transformers Pipeline. The Pipeline was using internal heuristics to map logits to labels (via softmax + sorting) and other approximations (rounding, label normalization, etc.).