[red-knot] detect unreachable attribute assignments

[mtshiba]

Summary

This PR closes #15967.

Attribute assignments that are statically known to be unreachable are excluded from consideration for implicit instance attribute type inference. If none of the assignments are found to be reachable, an unresolved-attribute error is reported.

Test Plan

A test case marked as TODO now work as intended, and new test cases have been added.

[github-actions]

mypy_primer results

No ecosystem changes detected ✅

[mtshiba]

While trying to solve the issue, I noticed that the following code does not pass:

class C:
    def __init__(self):
        def closure():
            self.a: str = 1
        closure()
        [... for self.b in range(1)]
        class D:
            self.c = 1

# should be OK
C().a
C().b
C().c

It appears that the SemanticIndexBuilder::register_attribute_assignment does not take inner scopes into account.
I'm trying to fix this, but this may be an issue that should be resolved in another PR.

[codspeed-hq]

CodSpeed Performance Report

Merging #16852 will not alter performance

_{Comparing mtshiba:unreachable-attribute-assignments (e52f9d6) with main (3aa3ee8)}

Summary

✅ 32 untouched benchmarks

[mtshiba]

The commit I just made completely removes SemanticIndex::attribute_assignments/AttributeAssignment.
Attribute assignments are now stored in SemanticIndex::all_definitions.

I noticed that we can also track attribute assignments in comprehension/named/augmented assignments as well, so I will be working on that.

[carljm]

Haven't finished my review but I have to head out, so submitting the comments I have so far. Thanks for your work on this!

Given the false positives observed, I'm not sure we should keep the possibly-unbound-attribute checks, but I'm still thinking about whether there is a satisfactory way to address the false positive cases. It may have some overlap with https://github.com/astral-sh/ruff/issues/15777

[mtshiba]

Support for attribute assignments in comprehension/aug-assign/named-expression is incomplete, but considering that it is still not implemented in the main branch and that it was not originally planned in this PR, this may be work that should be done in another PR.

I believe that the implementation of the feature originally planned in this PR is almost complete and ready for final review.

[carljm]

This looks pretty good to me! It turns ClassLiteralType::implicit_instance_attribute into even more of a monster method, but I think I'm ok with that for now: it has a clear responsibility, and the logic in it is well encapsulated; if we need to break it up for better reuse in future, we can do that.

@sharkdp if you have a chance, I'd like to make sure this looks good to you as well.

[sharkdp]

This looks great — thank you very much. I have just one minor comment, which I can also resolve myself before merging this.