<!--
Thank you for contributing to Ruff! To help us out with reviewing,
please consider the following:

- Does this pull request include a summary of the change? (See below.)
- Does this pull request include a descriptive title?
- Does this pull request include references to any relevant issues?
-->

## Summary

In #7968, I introduced a
regression whereby we started to treat imports used _only_ in type
annotation bounds (with `__future__` annotations) as unused.

The root of the issue is that I started using `visit_annotation` for
these bounds. So we'd queue up the bound in the list of deferred type
parameters, then when visiting, we'd further queue it up in the list of
deferred type annotations... Which we'd then never visit, since deferred
type annotations are visited _before_ deferred type parameters.

Anyway, the better solution here is to use a dedicated flag for these,
since they have slightly different behavior than type annotations.

I've also fixed what I _think_ is a bug whereby we previously failed to
resolve `Callable` in:

```python
type RecordCallback[R: Record] = Callable[[R], None]

from collections.abc import Callable
```

IIUC, the values in type aliases should be evaluated lazily, like type
parameters.

Closes #8017.

## Test Plan

`cargo test`