While rebasing PR #871 i noticed that our concatenate operator has strange default values for dims:

Here's backend/numpy_backend.py:

    @classmethod
    def concatenate(cls, arrays, dim=1):
        return cls._np.stack(arrays, axis=dim)

Here's backend/tensorflow_backend.py:

    @staticmethod
    def concatenate(arrays, dim=1):
        return tf.stack(arrays, axis=dim)

Here's backend/torch_backend.py:

   @staticmethod
    def concatenate(arrays, dim=2):
        return torch.stack(arrays, dim=dim)

I'm a bit confused by this dim=2 default value. And not too happy that the kwargs is called dim instead of axis, which would be more compliant with the Python Array API.

Our decision here will affect the semantics of #902

@eickenberg thoughts?