feat: add boolean cases for CPU kernels

[ManasviGoyal]

• Removed the following CPU Kernels and replaced their implementation with Numpy/Cupy -

  • awkward_unique
  • awkward_unique_copy
  • awkward_NumpyArray_fill

• Added boolean cases for same kernels

• Fixed and added some CUDA kernels

• Tests!

[ManasviGoyal]

@jpivarski I had removed awkward_unique and awkward_unique_copy and replaced it with nplike.unique_values which makes the CUDA implementation simple too. But I see that it is failing for ubuntu minimal test as the equal_nan argument requires numpy >= 1.19.0. Should I go back to using the kernels and add a separate case for the CUDA kernels or would it be possible to bump up the minimum version?

[jpivarski]

As usual, replacing kernels with NumPy/CuPy equivalents is a good thing to do, and I don't want to return to the custom kernel now that this mostly works.

One error, with the BitMaskedArray, might be because the whole content is being checked for uniqueness, rather than just the part up to the BitMaskedArray's length.

For the missing equal_nan argument in old versions of NumPy, you can implement it in the nplike layer, similar to the way that byteorder has been added to packbits for old NumPy versions. In the case of equal_nan, you would just need to check isnan for floating point types and do a logical_and of equality and isnan. It's not easy to check NumPy to see whether an argument exists (other than by version), but I don't think it would be bad to use this custom logic on all floating point types.

[jpivarski]

I verify that the CUDA tests all pass. All that's left is to fix the old-NumPy case (np.unique not having an equal_nan argument), and then this will be ready to merge.

[jpivarski]

This looks like it's done. If so, please merge it. Thanks!