array_api_compat.torch.take with 0-d indices? #460

array_api_compat.torch.take with 0-d indices produces a result with the same number of dimensions as the input.

from array_api_compat import torch as xp
x = xp.asarray([1])
xp.take(x, xp.asarray(0))
# tensor([1])

For non-wrapped torch.take (and all other array libraries) the output has one fewer dimension than the input.

import torch
x = torch.tensor([1])
torch.take(x, torch.asarray(0))
tensor(1)

from array_api_compat import numpy as xp
x = xp.asarray([1])
xp.take(x, xp.asarray(0))
# np.int64(1)

Technically, the standard of take requires indices to have exactly one dimension.
However, it also says:

Conceptually, take(x, indices, axis=3) is equivalent to x[:,:,:,indices,...];...

This is up for discussion in data-apis/array-api#1010.

But I'd think that either all libraries should follow the letter of the standard as closely as possible in this situation ("The output array must have the same rank (i.e., number of dimensions) as x") or array_api_compat.torch.take should be consistent with torch.take.