Fixes #1

Implementation is relatively straightforward, although I don't handle all edge cases yet. I manually checked this fixes few dozen tests in mypy. Couple notes:

• It looks like for some reason Ruff lexer doesn't have any public function that exposes full token stream. I only found a function returning token kinds. So for now we are forced to call parse() on the original function type comment just to get the token stream and split comment into argument types and return type.
• There are some .clone() calls on the parsed comments. I am not a Rust performance expert, but it looks like this is inevitable since we need to mutate them (relocate to meaningful locations, so that they can be used as context for type errors in mypy).