Type chunkmanagers

[Illviljan]

Use chunkedduckarray and duckarray in chunkmanagers now that it is used in namedarray.

Split from #8933
Requires #9226, #9225

[headtr1ck]

You should still use TypeVars and not the protocol directly.

[Illviljan]

I don't think that works (yet, waiting on Higher-Kinded TypeVars). Using T_DuckArray will indeed retain the np.ndarray typing (good) but also the dtype and shape typing (bad/wrong).

We don't really care (in theory) if it is ndarray or cupy array, just that they pass the minimal requirements for a duckarray. But we do care that dtype and shape is correct after calculation.

duckarray is actually duckarray[_ShapeType, _DType] where the dtype and shape are TypeVars, in same fashion as numpy see #8294. I am intentionally lazy with this in non-namedarray files though to avoid the PRs touching every file.

[Illviljan]

pd.index is apparently not a "duckarray":

    def test_pd_index_duckarray(self) -> None:
        import pandas as pd

        a: duckarray[Any, Any] = pd.Index([])
        check_duck_array_typevar(a)

xarray/tests/test_namedarray.py: note: In member "test_pd_index_duckarray" of class "TestNamedArray":
xarray/tests/test_namedarray.py:366: error: Incompatible types in assignment (expression has type "Index[int]", variable has type "_arrayfunction[Any, Any] | _arrayapi[Any, Any]")  [assignment]

FAILED xarray/tests/test_namedarray.py::TestNamedArray::test_pd_index_duckarray - TypeError: a (<class 'pandas.core.indexes.base.Index'>) is not a valid _arrayfunction or _arrayapi