Refactor semantic analyzer

[JukkaL]

The current semantic analyzer is hard to understand and modify. Here some ideas for how to improve things:

• Create separate modules for the three different passes (say, mypy.analyzer[123]).
• Don't pass a SemanticAnalyzer instance to the first and third passes. Instead, move any shared functionality to a utility class (or functions), and put them in a separate module such as mypy.analyzeutil.
• Rename the passes to, say, SemanticAnalyzerPass[123].

[emmatyping]

Should we close #3459 in favor of this? Or close this in favor of that?

[ilevkivskyi]

We can keep #3459 to track documenting part, which is not mentioned here, but I leave this up to @JukkaL to decide.

[elazarg]

Also

• We can try (again) to extract the creation of namedtuple, typeddict etc. to a separate module
• The analysis of a class can follow the interpreter: first analyzing the body, then passing the results to a "meta" function that constructs the class object. This will be cleaner, and a good plugin point (e.g. for special classes)

[emmatyping]

I think this issue could use an update since there has been a lot of progress on this :) I believe we still want to not pass the second pass' instance to the first pass, but are there other concrete steps left in this issue?

[JukkaL]

Here's what has been done:

• Semantic analysis passes have been moved to different modules and they have documentation.
• Named tuple, TypedDict, NewType and call-based Enum have been moved to new modules.

The main missing thing is getting rid of direct references between passes.

[JukkaL]

Closed in favor of #6204, which has a wider scope.