more fixes for NumPy2

[mattip]

Continuation of #6076 to make tests pass on NumPy2 and still support NumPy1.

The complex changes incorporate the fix for numpy/numpy#26029. Unfortunately on C++, NumPy no longer supports a[1].real += 1. It still works in C, but not in C++. This is documented in the migration guide (last paragraph in that section).

[mattip]

One CI run failed to update with apt, and a windows/python3.13 one failed in test_initvar_is_specified.

[scoder]

LGTM. Thanks.

[da-woods]

3.0.x commit 4f3c1ba

[scoder]

I now see C compile failures in 3.0.x, both in C and C++:
https://github.com/cython/cython/actions/runs/8401443347/job/23009827687

In file included from numpy_test.c:32:
numpy_test.c:3882:42: error: request for member ‘imag’ in something not a structure or union
   {&__Pyx_TypeInfo_long__double, "imag", offsetof(__pyx_t_5numpy_clongdouble_t, imag)},
                                          ^~~~~~~~
numpy_test.c:3882:42: warning: initialization of ‘long unsigned int’ from ‘__Pyx_StructField *’ {aka ‘struct __Pyx_StructField_ *’} makes integer from pointer without a cast [-Wint-conversion]
   {&__Pyx_TypeInfo_long__double, "imag", offsetof(__pyx_t_5numpy_clongdouble_t, imag)},
                                          ^~~~~~~~
numpy_test.c:3882:42: note: (near initialization for ‘__Pyx_StructFields_nn___pyx_t_5numpy_clongdouble_t[1].offset’)
In file included from numpy_test.c:32:
numpy_test.c:3882:42: error: initializer element is not constant
   {&__Pyx_TypeInfo_long__double, "imag", offsetof(__pyx_t_5numpy_clongdouble_t, imag)},
                                          ^~~~~~~~
numpy_test.c:3882:42: note: (near initialization for ‘__Pyx_StructFields_nn___pyx_t_5numpy_clongdouble_t[1].offset’)

numpy_test.cpp:3896:81: error: ‘__pyx_t_5numpy_clongdouble_t’ {aka ‘struct npy_clongdouble’} has no member named ‘real’; did you mean ‘_Val’?
   {&__Pyx_TypeInfo_long__double, "real", offsetof(__pyx_t_5numpy_clongdouble_t, real)},
                                                                                 ^~~~
numpy_test.cpp:3897:81: error: ‘__pyx_t_5numpy_clongdouble_t’ {aka ‘struct npy_clongdouble’} has no member named ‘imag’
   {&__Pyx_TypeInfo_long__double, "imag", offsetof(__pyx_t_5numpy_clongdouble_t, imag)},
                                                                                 ^~~~

[mattip]

Huh. I tested this with python3.11, maybe something changed with python3.12? Investigating...

[mattip]

Ahh, of course. This needs latest numpy HEAD to pull in the changes made in numpy/numpy#26080 after the release of 2.0.0b1, so it will fail until the release of 2.0.0rc1. So what should be done here?

• There are newer "nightly" wheels available, they can be used with pip install -i https://pypi.anaconda.org/scientific-python-nightly-wheels/simple numpy==2.1.0.dev0. Should I submit a PR to use that until a better release is on PyPI?
• Disable the test as long as NumPy version == 2.0.0b1
• Live with the failure until the rc1 release, which could happen in a week or twp

[mattip]

Another option is to pin 3.0.x to use numpy<2 on 3.12 like master, at least until after the rc1 release.

cython/test-requirements-312.txt

Line 1 in 3113c7f

numpy<2

[scoder]

I can live with the failures for the time being, as long as we know where they come from and that they'll go away in the not so distant future.