Fix a mathematical bug

[tushuhei]

This change fixes the mathematical bug in the parse method, which eventually removes the necessity of the thres parameter entirely. This also fixes the deviation between the reported metrics during model training and actual quality of the results provided by the parser.

This PR includes:

• a small fix to remove line breaks from the training data, which will make the resulting parser robust in processing punctuations that often come to the end of sentences.
• retrained model files based on the change above.
• updated parser implementation with correct score calculation logic and no thres parameter.

⚠️ Breaking change
thres won't be available in the parse method and the CLI options any more. Please fix your program if it's relying on the thres parameter.

[kojiishi]

lgtm

[tushuhei]

Thanks for reviewing!