Fix loop convergence with redefinitions

[ilevkivskyi]

Currently, we reset the variable type each time we visit the initial definition (textually first assignment). This can cause problems with --allow-redefinition-new:

• Not all branches will be explored (see added test). This is a correctness problem.
• Many loops will become runaway (will never converge to a fixed point) while they actually can converge. Currently this is stopped by manual cut-off. This is a performance problem.

The only non-trivial part here is handling of this pattern:

for key in ("key_a", "key_b"):
    some_typed_dict[key]

It is special-cased in regular mode, and this special-casing used to work accidentally with redefinition allowed. Now I simply make the same special-casing apply (since for loops can never be runaway on index).

[ilevkivskyi]

Repro for meson difference:

[case testNewRedefineAnyOrEmptyUnpackInLoop]
# flags: --allow-redefinition-new --local-partial-types
from typing import Any

def foo() -> None:
    while int():
        data: Any
        x, y = data or ([], [])
        reveal_type(x)
[builtins fixtures/isinstancelist.pyi]

Currently we silently infer Any | list[Never], with this PR we give the error and infer Any | list[Any]. TBH I am not sure this is a regression. Also there is a chance this will be fixed by #20851 Actually, I will probably just merge that one, and try re-basing this PR.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅