Previously the the input json was directly mapped into the output
strings, through different formatters. This changes it so we instead
first build a proper AST of python objects, which is later mapped to
output strings (still through formatters).

This gives us much better type information, and will give way to much
more advanced formatters in the future.

This includes some regressions:
* Docstrings aren't handled as well
* The code lacks any form of hyperlinks.