Refactor `apply_vocabulary`

[yurytsoy]

The current implementation is naive and has (at least) the following problems:

• Addresses each token separately with multiple passes over the original template
• Might benefit from incapsulation of the conjugations and inflections
• Takes way too many lines than needed without obvious benefits for readability.

These can be improved in few ways:

• Add regexps instead of plain matching of substrings
• Move inflections into the dictionary
• Make a mapping between token and function for generation of a word.