Updated mk_alloc to support Numba-Dpex compute follows data.

[mingjie-intel]

Change is made in mk_alloc to support Numba-Dpex compute follows data. There will be no affect to Numba main function.

[DrTodd13]

Please add this allocate support in the appropriate place in the Numba developer documentation.

[diptorupd]

Some more motivating context behind these changes. The mk_alloc function is used inside the Numba parfor machinery to create a new empty NumPy array (by calling numpy.empty) whenever a temporary array is needed to allocate the left-hand-side sub-expression of a parfor node.

In our work on using parfors to support other NumPy-like libraries we want to change the actual library call used to allocate the empty array. These changes extend the mk_alloc function to first look if the array type provides an __allocate__ function and use that instead of the default logic in the original function.

@DrTodd13 can you point me to the file where the documentation should be updated?

[DrTodd13]

@diptorupd @sklam @stuartarchibald I don't know if the documentation belongs in numba/docs/source/extending or numba/docs/source/developer. Unbelievably, we don't have a ParallelAccelerator page in the developer docs which was my first idea. I'm soliciting ideas from Siu and Stuart as to where this should be documented.

[sklam]

/azp run

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[sklam]

@diptorupd @sklam @stuartarchibald I don't know if the documentation belongs in numba/docs/source/extending or numba/docs/source/developer. Unbelievably, we don't have a ParallelAccelerator page in the developer docs which was my first idea. I'm soliciting ideas from Siu and Stuart as to where this should be documented.

I think this is ParallelAccelerator specific enough to have its own page. I think we can add that ParallelAccelerator page in a different PR.