fix generation with constraints for decoder models

[Ingvarstep]

It was fixed generation with constraints for decoder-based models:

• We truncate sent sequence of tokens removing all tokens that doesn't belong to label;
• It was changed mechanism of adding new pad_token, instead of increasing embedding size we initialize pad token as eos token;
• We artificially add pad token to sent sequence that belongs to label, because this is how we encoded it in a trie;
• It was truncated the output string generated by decoder removing all prompt text.