idxmax with multiple dimensions

[gcaria]

• Closes idxmax with multiple dimensions #10112
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst

[max-sixty]

Looks good @gcaria !

Would it be possible to add a couple tests?

[dcherian]

we should have some type stability here so

idmax(dim) -> array; idxmax((dim,)) -> tuple[array]; idxmax((dim0, dim1, ...)) -> tuple[array, ...]

[gcaria]

I have changed the code to match the behavior of DataArray.arg* which returns a dict for both idx*((dim,)) and idx*((dim0, dim1, ...))

Does that seem sensible?

Currently navigating the existing tests for arg* and multiple dims

[gcaria]

While playing with the tests for arrays that contain nans, I found this, which doesn't seem to make sense (running on main):

import xarray as xr
import numpy as np

x = np.array([[ 2.,  1.,  2.,  0., -2., -4.,  2.],
       [-4., np.nan,  2., np.nan, -2., -4.,  2.],
       [np.nan, np.nan, np.nan, np.nan, np.nan, np.nan, np.nan]])

ar = xr.DataArray(x,
                dims=["y", "x"],
                coords={
                "x": np.arange(x.shape[1]) * 4,
                "y": 1 - np.arange(x.shape[0])
},)

ar.idxmin('x', skipna=False) # ---> array([20,  4,  0])

Shouldn't the result be array([20, nan, nan]) (or error, see below)?
With the default skipna=True I do get array([20., 0., nan])

Interestingly ar.argmin('x') instead raises ValueError: All-NaN slice encountered, which I'd expect to show up also for idxmin.