feat: make numpy.h compatible with both NumPy 1.x and 2.x (#5050)

* API: Make `numpy.h` compatible with both NumPy 1.x and 2.x

* TST: Update numpy dtype flags test to not covert flags to char

* API: Add `numpy2.h` instead and make `numpy.h` safe

This means that users of `numpy.h` cannot be broken, but need to
update to `numpy2.h` if they want to compile for NumPy 2.

Using Macros simply and didn't bother to try to remove unnecessary
code paths.

* API: Rather than `numpy2.h` use a define for the user.

* Thread `PYBIND11_NUMPY2_SUPPORT` through things and try to adept test matrix

* Small fixups (shouldn't matter)?

* Fixup.  Does upgrading scipy help?  (it shouldn't?)

(Some other small fixup)

* Use NumPy 2 nightlies for ubuntu-latest job also

* BUG: Fix numpy.bool check

* TST: Fix complexwarning

* BUG: Fix the fact that only the 50 slot is filled with the copy alias

(There were 3 functions all doing the same, only this slot survived 2.x)

* TST: One more test tweak

* TST: Use "long" name for long, since it changed on windows

* TST: Apparently we didn't always have ulong, so just use `L`

* TST: Enforce dtype='l' for test as default isn't long anymore on windows

* Rename macro and invert logic to PYBIND11_NUMPY_1_ONLY

* PYBIND11_INTERNAL_NUMPY_1_ONLY_DETECTED

* Test and code comment expansion

* CI: Use pre-releases of numpy/scipy from pip via explicit version

* CI: NumPy 2 only available on almalinux (as it is Python >=3.9)

* MAINT: Match name more exactly and adopt error phrasing

* MAINT: Pushed early, move helper to be private member

* fix error message compilation when using NumPy 1.x-only backcompat

* silence name shadowing warning

* chore: minor optimization

Signed-off-by: Henry Schreiner <henryschreineriii@gmail.com>

---------

Signed-off-by: Henry Schreiner <henryschreineriii@gmail.com>
Co-authored-by: Ralf W. Grosse-Kunstleve <rwgk@google.com>
Co-authored-by: Henry Schreiner <henryschreineriii@gmail.com>

tests/test_numpy_array.py

@@ -536,7 +536,12 @@ def test_format_descriptors_for_floating_point_types(test_func):
 @pytest.mark.parametrize("contiguity", [None, "C", "F"])
 @pytest.mark.parametrize("noconvert", [False, True])
 @pytest.mark.filterwarnings(
-    "ignore:Casting complex values to real discards the imaginary part:numpy.ComplexWarning"
+    "ignore:Casting complex values to real discards the imaginary part:"
+    + (
+        "numpy.exceptions.ComplexWarning"
+        if hasattr(np, "exceptions")
+        else "numpy.ComplexWarning"
+    )
 )
 def test_argument_conversions(forcecast, contiguity, noconvert):
     function_name = "accept_double"
@@ -583,7 +588,8 @@ def test_argument_conversions(forcecast, contiguity, noconvert):
 def test_dtype_refcount_leak():
     from sys import getrefcount
 
-    dtype = np.dtype(np.float_)
+    # Was np.float_ but that alias for float64 was removed in NumPy 2.
+    dtype = np.dtype(np.float64)
     a = np.array([1], dtype=dtype)
     before = getrefcount(dtype)
     m.ndim(a)