Fix ufunc docstrings showing the C macro name

[antonwolfy]

Summary

Fixes the __doc__ of every mkl_umath ufunc, which ended with a literal C macro name instead of the docstring text:

>>> mkl_umath.sin.__doc__
"sin(x, /, out=None, *, where=True, casting='same_kind', order='K', dtype=None, subok=True[, signature])\n\nDOC_NUMPY__CORE_UMATH_SIN"

All 58 public ufuncs were affected, in every release since 0.3.0. NumPy's own ufuncs are not affected by patching, since patch_numpy_umath() swaps loops, not ufunc objects.

Root cause

gh-65 aligned generate_umath.py with NumPy 2.2's docstring scheme: docstrings.get() returns the name of a DOC_* C macro, defined in a _umath_doc_generated.h header produced by generate_umath_doc.py. Two pieces of that scheme were not carried over:

• the generator kept emitting the doc argument quoted ("{doc}"), a leftover from the previous scheme where it held the docstring text itself, so the C code got the string "DOC_NUMPY__CORE_UMATH_SIN";
• generate_umath_doc.py was ported but never invoked by the build, so the header defining the macros was neither generated nor included.

Fix

• meson.build: add a custom_target that generates _umath_doc_generated.h (rerun when ufunc_docstrings_numpy{1,2}.py change) and add it to the _ufuncs sources.
• generate_umath.py: emit the macro unquoted and #include "_umath_doc_generated.h" in the generated code, matching upstream NumPy. The obsolete "difference Transition from numpy.distutils to scikit-build #3" note in the file header is removed.
• test_basic.py: add test_docstring, checking each ufunc docstring contains no DOC_ and has its Parameters section.

Every DOC_* macro referenced by the generated code is defined by the generated header, and both ufunc_docstrings_numpy1.py and ufunc_docstrings_numpy2.py use the numpy._core.umath keys the generator expects. The header splits docstrings into adjacent literals at newlines, as NumPy does, so MSVC's string literal length limit is not hit.

Testing

Linux, icx, Python 3.13, NumPy 2.5.4: full test suite passes (450 passed, 1 skipped). The new test fails for all 58 ufuncs on a build from main. Windows, clang and NumPy 1.x builds are covered by CI.

[jharlow-intel]

Thanks, this is a nice fix! Fixing it in the generator means every ufunc gets a real docstring under both NumPy 1 and 2, and the new parametrized test will keep it that way. Just one small changelog note inline.