Require NumPy >= 2 and remove the NumPy 1 code paths

[antonwolfy]

mkl_umath picks its NumPy 1 or NumPy 2 loops at compile time (USE_NUMPY_2), but declares numpy >=1.26.4 at runtime. Every CI job and every published package since 0.3.0 is built against NumPy 2, so it can be installed with NumPy 1.x and then misbehaves. This PR requires NumPy 2 everywhere and removes the NumPy 1 code that nothing builds or tests anymore.

Problem

• With the published 0.5.0 wheel (cp312, Linux) and NumPy 1.26.4, 4 tests fail: complex sign (F->F, D->D) in test_mkl_umath and test_fall_back_umath. After patch_numpy_umath(), np.sign(3+4j) returns 0.6+0.8j instead of NumPy 1's 1+0j. With NumPy 2.0.2 all tests pass.
• On Windows, NumPy 1.x's ldexp only has int and long (32-bit there) loops, so patch_numpy_umath() finds no NumPy loops matching the int64 ldexp loops and raises RuntimeError (from reading _patch_numpy.pyx, not run on Windows).
• All wheels from 0.4.0 to 0.5.0 (identical on PyPI and the Intel index) are built against NumPy 2 and allow NumPy 1.26.x; the NumPy 2 builds of 0.2.0–0.3.1 required NumPy 2. The lower bound dropped in task: numpy-base toggle for meta.yaml #181, which replaced pin_compatible('numpy-base') with numpy >=1.26.4 in conda-recipe/meta.yaml.

Changes

Require NumPy 2

• pyproject.toml: dependencies = ["numpy >=2"] and numpy>=2 in [build-system] requires.
• conda-recipe/meta.yaml: run numpy >=2 (was >=1.26.4). The published wheels take their requirements from this recipe's run dependencies.
• conda-recipe/conda_build_config.yaml: numpy: '2' (was '1.26.4'), so a local conda build conda-recipe without --numpy builds against NumPy 2. CI always passes --numpy.

Remove the NumPy 1 code paths

• mkl_umath_loops.{c,h}.src: drop the USE_NUMPY_2 switch, the NumPy 1 complex sign loop and the ldexp_long loops, and keep the NumPy 2 code unconditionally. The header now stops the build with #error "mkl_umath must be built against NumPy >= 2.0" when compiled against NumPy 1 headers.
• generate_umath.py, generate_umath_doc.py: drop the NumPy version branches and the now unused numpy imports.
• Delete ufunc_docstrings_numpy1.py and its .flake8 entry, and update the AGENTS.md references.
• CHANGELOG entries under [dev].

Testing

• Built with icx against NumPy 2.5.4 without warnings. The wheel has Requires-Dist: numpy>=2 and exports only the ldexp_int64 loops.
• pytest --pyargs mkl_umath on py3.13 with NumPy 2.1.3 (the oldest NumPy for py3.13) and 2.5.4: 392 passed, 1 skipped.
• The header compiles with NumPy 2.5.4 headers and stops at the #error with NumPy 1.26.4 headers.

Notes

• conda-recipe-cf gets the same numpy >=2 run requirement and numpy: '2' default in Align conda-recipe-cf and its CI with the conda-forge feedstock #287. The comment Align conda-recipe-cf and its CI with the conda-forge feedstock #287 adds there mentions USE_NUMPY_2, so whichever PR lands second should reword it. Until Align conda-recipe-cf and its CI with the conda-forge feedstock #287 is merged, a local conda build conda-recipe-cf without --numpy stops at the new #error; CI is not affected.
• The conda-forge feedstock is updated in Require NumPy 2 at runtime conda-forge/mkl_umath-feedstock#9.
• Published wheels cannot be changed on PyPI, so this applies from the next release.
• Unrelated, found while removing ufunc_docstrings_numpy1.py: generate_umath_doc.py is not wired into meson.build, so the ufunc docstrings end with the literal macro name (mkl_umath._ufuncs.add.__doc__ ends with DOC_NUMPY__CORE_UMATH_ADD) and ufunc_docstrings_numpy2.py is not used either. Left for a separate PR.