implement @no_type_check for class defs

[lunixbochs]

Fixes #607 (which is incorrectly titled).

This adds TypeInfo.is_no_type_check, and enables TypeInfo.fallback_to_any for classes decorated with @no_type_check, which is checked in a couple of places to abort the type checker early and stub method types as Any.

I'm not sure confident about my use of decorator.name in ('no_type_check', 'typing.no_type_check' in semanal.py. I see RefName.fullname = 'typing.no_type_check' was used elsewhere but it wasn't a RefName in the AST during semanal and I didn't look too far into that.

This is my current test:

from typing import no_type_check, Any

@no_type_check
class TestClass:
    bar: {'x': 1} = 1

    def meth1(): pass
    def meth2(a: {'x': 1}) -> {'x': 1}:
        def garbage_inner(a: {'x': 1}):
            pass

    def test(a: int): pass

TestClass().test(1)
TestClass().doesntexist()

class Test(TestClass):
    def test2(self, a: int): pass

t = Test()
t.meth1()
t.doesntexist2()
t.test(1)
t.test2(1)

x: int = TestClass().bar

With @no_type_check, mypy throws a Missing return statement for the nested garbage_inner definition, but suppresses the rest of the type errors for this class and subclass.

Without @no_type_check, we get this nice pile of errors:

repro.py:4: error: Invalid type comment or annotation
repro.py:6: error: Method must have at least one argument
repro.py:7: error: Invalid type comment or annotation
repro.py:8: error: Invalid type comment or annotation
repro.py:11: error: Self argument missing for a non-static method (or an invalid type for self)
repro.py:13: error: Too many arguments for "test" of "TestClass"
repro.py:14: error: "TestClass" has no attribute "doesntexist"
repro.py:21: error: "Test" has no attribute "doesntexist2"
repro.py:22: error: Too many arguments for "test" of "TestClass"

TODO (I'd like advice from core contributors on these):

• I think I should write in-tree tests for this, but I could use guidance on that.
• confirm my decorator.name check is correct.
• is it possible to suppress the no return error?
• am I missing any additional obvious class type behavior in my test that should be suppressed/stubbed by the decorator?

[lunixbochs]

Can a maintainer look at this / comment?

[ilevkivskyi]

Few comments:

• This is still a draft, please make it non-draft and resolve merge conflicts if you want to get a review
• Your decorator.name check is wrong, you should check for a RefExpr and its fullname instead.
• I don't think we need a new flag for this. I think fallback_to_any plus replicating effect of @no_type_check for every method should be enough.
• Please add tests for this, I think https://github.com/python/mypy/blob/master/test-data/unit/check-classes.test is the best place.

[lunixbochs]

• At the point of SemanticAnalyzer.analyze_class, all defn.decorators are still of type NameExpr. How/where can I get RefExpr decorators instead at that point?
• I used the special flag because I needed to propagate the behavior to semanal3 and analyze_instance_member_access, which fallback_to_any isn't expressive enough to do. I don't see how to propagate the behavior to methods at all at the point where we're just skipping analyze_class.

I'm also now getting fallback can't be filled out until semanal I guess due to skipping analysis with the new semantic analyzer.

[msullivan]

I'm going to close this because it is against the old semantic analyzer which is totally gone now. Feel free to redo this against the current code base and I'll review it.