Allow `ndarray_check` on objects of arbitrary modules

[njroussel]

This PR changes ndarray_check such that it returns True if one of the following conditions is met:

• The given object has a __dlpack__ attribute
• The given object supports the CPython buffer interface
• The given object is an object from NumPy (implicit through buffer protocol), PyTorch, XLA, TensorFlow

With this change, ndarray_check more closely matches the logic of supported imports through ndarray_import.

I've added a straightforward test which checks this new behavior on an object with a __dlpack__ attribute.

[wjakob]

Thanks!