[FEA] Support for `__dlpack_capi__`

[gmarkall]

As part of a general request for all array / data containers in CUDA and Python to support the __dlpack_capi__ API, there is a request for Numba-CUDA to implement this too.

The most efficient / fastest implementation may be hard to do with Numba-CUDA device arrays in their present form - the shape, strides, data pointer, etc. are stored on a Python subclass of the base device array class. The base device array is a plain PyObject struct, as outlined:

numba-cuda/numba_cuda/numba/cuda/cext/_devicearray.cpp

Lines 17 to 24 in a194f90

	/* DeviceArray PyObject implementation. Note that adding more members here is
	* presently prohibited because mapped and managed arrays derive from both
	* DeviceArray and NumPy's ndarray, which is also a C extension class - the
	* layout of the object cannot be resolved if this class also has members beyond
	* PyObject_HEAD. */
	class DeviceArray {
	PyObject_HEAD
	};

Therefore, a __dlpack_capi__ implementation would need to go back to CPython to get the necessary data, rather than it being stored in the DeviceArray struct. There may be some refactoring possible to allow us to extend this structure to retain the necessary information at the C layer, but some exploration is needed to discover the possible / necessary changes.

There is likely still some performance gain from implementing this API, by bypassing the construction of new Python objects compared to implementing e.g. __dlpack__.

I am unsure whether the request is also that kernel launches should operate of objects exposing __dlpack_capi__, but it seems that this may provide a path for faster launches.

cc @oleksandr-pavlyk

[leofang]

I left a comment on DLPack here: #153 (comment)

[leofang]

@oleksandr-pavlyk Question: With #471 + #153 in mind, is it still necessary for Numba-cuda to add __dlpack_capi__ support, or would this be better done in cuda.core and/or CuPy?

[oleksandr-pavlyk]

I think it should be done in actively containers. In DeviceArray is on the way out, we don't need to extend it.

We should transition away from using DeviceArray in cuda.compute test suite then.