Support python 3.15 lazy imports #13559

The biggest question I have with lazy imports is how they will interact with relationships in model definitions with cyclic dependencies when running introspection on Mapped annotations. A lazy import will behave differently from an import inside a type checking block, since the lazy import will trigger the import and then potentially a circular dependency error, the import inside a type checking block will trigger a NameError instead, since the name is never defined.

If get_type_hints/get_annotations can recover from both errors in the same way without side effects with Format.FORWARDREF, then there's nothing we need to do, but if there is an actual runtime difference between those two error conditions, that causes significant downsides for the lazy import version, then it seems important that we try to make them somehow both work.

The biggest worry I have here is that the lazy import either becomes permanently broken in that module, or that half a dozen broken imports are attempted at module execution time, only to fail and later having to start over, slowing down initial load times, rather than improving them.

Python tooling will almost certainly start pushing people away from if TYPE_CHECKING blocks and towards using lazy imports instead and if that starts breaking their programs relying on SQLAlchemy, they will become grumpy.