Moving TestDeviceBuffer to pylibraft.common.device_ndarray

[cjnolet]

This allows the pylibraft APIs to process outputs in place or allocate and return new device memory for outputs in a way which benefits from __cuda_array_interface__ compliance. Ultimatley, it allows us to stay decoupled from libraries like cupy while still affording the user the benefits of not having to explicitly allocate outputs up front.

Closes #1002

[benfred]

I really like this change - I think its a solid improvement to the usability of the python bindings, while still not requiring cupy

[cjnolet]

Thanks @benfred!

I'm thinking we should also provide a factory function that accepts the shape, order, and dtype and creates an empty numpy array (internally) to construct the device_ndarray(). Maybe something the caller can invoke like this:

dev_ndarr = device_ndarray.empty((50, 10), dtype=np.float32, order="C")

Obviously, there's a drawback to using the device_ndarray since it requires a host/device copy. However, I think that's okay because we are giving the option of ease of use (eg have the function automatically return results) and speed (eg users can still explicitly pass in the outputs to keep everything on device and avoid using the device_ndarray)

[cjnolet]

JFYI, I added the pylibraft.device_ndarray.empty() staticmethod and updated its use in pairwise_distance.

[benfred]

this looks good! I see your point about the extra h2d memory copy w/ device_ndarray.empty - but I agree that users can always provide pre-allocated arrays to get around this

[cjnolet]

@gpucibot merge