CUDA: Add binary ufunc support

[Matt711]

Following from #8294, this PR adds support for binary ufuncs for the CUDA target (bitwise_and, bitwise_or, etc.). Currently there aren't any unit tests for np.left_shift and np.right_shift. I didn't add them because there was a comment here saying the tests were not added for the CPU target either.

[gmarkall]

From a first pass over this, I think this is looking good like the pattern for implementing ufuncs in CUDA is followed correctly - I have a suggestion and a comment on the diff for now, and will have another pass once I understand the responses.

[gmarkall]

Why do we not have implementations for unsigned types?

[Matt711]

Thanks! I added the missing types from np.<ufunc>.types except O which is a generic python object

[gmarkall]

gpuci run tests

[gmarkall]

gpuci run tests

[gmarkall]

Many thanks for the updates! A couple of quick questions / observations on the diff.

[gmarkall]

I think this is superflous, and just reassigns the same key in the dict - these two keys are the same:

$ python -c "import numpy as np; print(np.invert is np.bitwise_not)"
True

I think this can be removed.

[gmarkall]

gpuci run tests

[gmarkall]

Many thanks for the updates! This looks good now.

I think I'd like a buildfarm run for this, since it's the first set of new ufuncs we've added since the addition of the feature - maybe eventually these additions will seem "safe" / generally expected to work everywhere if they work in gpuCI, but I'm not confident there yet.

@stuartarchibald / @sklam Could this have a buildfarm run please?

[gmarkall]

I understand from an OOB message that numba_smoketest_cuda_yaml_205 ran and passed on this PR.