This PR adds support to call kernels in CUDA from v2 Awkward Arrays.

[findingmochi]

Tasks to complete:

• Implement copy_to for all array classes
• Implement to_backend and backend(*arrays)
• Write a minimal test using ak.num
• Call a kernel whose implementation is not in CUDA and see if it's being handled correctly.
  - [ ] Don't upload artifacts during CI testing.

[codecov]

Codecov Report

Merging #1276 (47c5e8b) into main (8e68200) will increase coverage by 1.27%.
The diff coverage is 74.82%.

Impacted Files	Coverage Δ
src/awkward/_v2/_connect/numba/arrayview.py	96.24% <ø> (ø)
src/awkward/_v2/_connect/numba/layout.py	87.01% <ø> (ø)
src/awkward/_v2/operations/convert/ak_from_cupy.py	25.00% <0.00%> (-50.00%)	⬇️
src/awkward/_v2/operations/structure/ak_isclose.py	100.00% <ø> (ø)
src/awkward/_v2/operations/convert/ak_to_cupy.py	8.19% <3.44%> (-66.81%)	⬇️
src/awkward/_v2/operations/describe/ak_backend.py	9.52% <9.52%> (ø)
src/awkward/_v2/contents/unmaskedarray.py	54.95% <25.00%> (-1.12%)	⬇️
...c/awkward/_v2/operations/describe/ak_to_backend.py	37.50% <37.50%> (ø)
src/awkward/_v2/contents/indexedarray.py	58.62% <38.23%> (-2.03%)	⬇️
src/awkward/_v2/operations/io/ak_to_json_file.py	47.05% <43.75%> (-27.95%)	⬇️
... and 49 more

[findingmochi]

The functions _getitem_next*(many _getitem_next_jagged) are passing mixed nplike because I think they are passing slices and slices are Numpy nplike through and through, but when the typetracers tests come in, these two nplikes get mixed and hence they raise the assertion errors introduced in Wednesday's refactoring.

For now, I have commented out the assertions and they should be easy to find by doing a fgrep -rl "# assert" src.

Also, now to has been replaced by raw and ak.to_cupy and ak.from_cupy functions have also have been implemented.

[jpivarski]

It's fantastic that this is working! On review, I found a few organizational things that, fixing them now, will save us trouble later. See the inline comments.

Thanks!