Update to numba-cuda 0.4.0 and allow numba 0.61.

[bdice]

Description

This PR updates to numba-cuda 0.4 and numba 0.61.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[bdice]

@gmarkall This PR reproduces the test failures we saw earlier. Could you take a peek?

Failures look like:

 FAILED test_unaops.py::test_scalar_unary_operations[False-bool-floor] - AssertionError: assert dtype('bool') == dtype('float32')
 +  where dtype('bool') = np.False_.dtype
 +  and   dtype('float32') = Scalar(0.0, dtype=float32).dtype

[gmarkall]

I think these are a result of numpy/numpy#26766 - I can't see the Numba / Numba-CUDA connection.

[vyasr]

What do we need to do here? Are we thinking that the problem is not in numba, but rather that numba was previously constraining the numpy version for us and that cudf needs to adapt now for the new numpy version?

[gmarkall]

Yeah, it's just that Numba was constraining the NumPy version. From NumPy 2.1 onwards, floor, ceil, and trunc will return integer-dtyped results when the input is an integer. I'm not sure of the right course of action because I don't have enough cuDF expertise, but... Looking at test_unaops, for example, I guess the options would be to:

• Modify the tests so that it accepts cuDF producing float results for these functions even when NumPy produces ints
• Modify unary ops so that they also produce int results where NumPy would have done.

I guess the answer might depend on whether pandas changes the dtype for operations like these when the NumPy version changes - I don't know which way is goes, though.

(If the above just makes no sense, please ignore me, I don't know what I'm talking about 🙂)

[vyasr]

@bdice I think the immediate question here is regarding the scope of this PR. Either you can add an explicit upper bound on numpy, unpin numba, open a separate issue for unpinning numpy, then get this PR merged, or you can try to resolve the numpy issues. I'd maybe suggest trying the latter quickly and if you get stuck then switch to the former, but really up to you.