Bradley-Terry result diverges from Crowd-Kit

Implementation of the Bradley–Terry model in Evalica diverges from the one used in Crowd-Kit in this example: [test.csv](https://github.com/user-attachments/files/17006815/test.csv). ``` # Crowd-Kit 1.4.1 c 0.378931 f 0.378931 a 0.146774 b 0.027966 d 0.027966 g 0.027966 e 0.011465 ``` ``` # Evalica 0.2.0 (with normalization) f 0.382009 c 0.382009 a 0.143738 b 0.027053 g 0.027053 d 0.027053 e 0.011086 ```