Consider having a `_typing.py` or a way to restructure common type aliases

[selmanozleyen]

refering to: #1076 (comment)

Maybe this? Or use the one from squidpy.read._utils instead of os.PathLike | str if that doesn’t lead to circular imports

    def from_yaml(cls, config_path: os.PathLike | str | None = None) -> DatasetRegistry:

Originally posted by @flying-sheep in #1076 (comment)

[selmanozleyen]

Would also help with putting SpatialResults in types module instead of graph

[flying-sheep]

OK, so my experience is that there are

1. Runtime usable objects that are also usable as types such as
   • Allowed = Literal["a", "b", …] used in get_args(Allowed)
   • SomeUnion = Foo | Bar used in isinstance(thing, SomeUnion)
   • Re-exports of optional dependencies used in isinstance checks ¹
2. type Foo = … aliases that are public and only used in type hints

All of this is a bit messy because sometimes you just want an alias for a type union to reuse in type hint code (but want it expanded in the docs) and sometimes you want to export it as a public linkable thing.

Footnotes

1. The ideal way to do this is like so, which makes isinstance checks work just like Sphinx auto-linking if you use it in a type hint:

   if TYPE_CHECKING or find_spec("dask"):
   	from dask.array import Array as DaskArray
   else:  # pragma: no cover
   	DaskArray = type("Array", (), dict(__module__="dask.array"))
   

   ↩

[selmanozleyen]

ok let me see the problems I come up with here. I also want to switch to TypedDicts like annbatch because result objects shouldnt be listed here