BUG: apply NEP 50 scalar rules in concatenate and choose, and make casting="no" match "equiv" - #32510

Closes #32491

Problem & Background

In ufuncs and np.copyto, passing casting="no" with a Python scalar whose target dtype differs from its default dtype (e.g. int64 for Python int, float64 for Python float) was previously accepted, while casting="equiv" raised a TypeError:

np.copyto(np.zeros(3, dtype="uint8"), 3, casting="no") # SUCCEEDED
np.copyto(np.zeros(3, dtype="uint8"), 3, casting="equiv") # FAILED with TypeError

Because "no" casting is strictly more restrictive than "equiv", "no" should never succeed where "equiv" fails.

Root Cause
In numpy/_core/src/multiarray/abstractdtypes.c, npy_update_operand_for_scalar specifically checked for casting == NPY_EQUIV_CASTING. When casting == NPY_NO_CASTING was passed, it fell through, recreated the temporary array with the target descriptor, and the subsequent cast-safety check compared identical dtypes.

Additionally:

np.concatenate(axis=None) and np.choose did not convert Python scalars using NEP 50 value bounds, causing out-of-bounds integers to silently wrap instead of raising OverflowError (e.g., np.concatenate((np.ones(2, "int8"), 300), axis=None) returned 44 for 300).
In abstractdtypes.c, npy_update_operand_if_pystr was only handling Python str, but the same scalar re-conversion logic was needed for all Python literals/scalars (NPY_ARRAY_WAS_PYTHON_LITERAL).
Changes Made

1. Multiarray Core C Engine

abstractdtypes.c:
Included convert_datatype.h for npy_casting_to_string.
Changed condition in npy_update_operand_for_scalar from casting == NPY_EQUIV_CASTING to casting <= NPY_EQUIV_CASTING, rejecting both "no" and "equiv" when the scalar's default dtype does not match the target. Formatted the exception message dynamically using npy_casting_to_string(casting).
Generalized npy_update_operand_if_pystr to npy_update_operand_if_pyscalar, handling NPY_ARRAY_WAS_PYTHON_LITERAL | NPY_ARRAY_WAS_PYTHON_STR and forwarding the operation's casting argument.

abstractdtypes.h:
Updated prototype for npy_update_operand_if_pyscalar.

convert_datatype.c:
Updated PyArray_ConvertToCommonType (used in np.choose) to call npy_update_operand_if_pyscalar(&mps[i], op, i, common_descr, NPY_SAFE_CASTING).

multiarraymodule.c:
Updated PyArray_ConcatenateFlattenedArrays (np.concatenate(..., axis=None)) to call npy_update_operand_if_pyscalar(&arrays[iarrays], op, iarrays, PyArray_DESCR(ret), casting).

2. Documentation & Release Notes
   Added Towncrier release notes:
   doc/release/upcoming_changes/32497.compatibility.rst
   doc/release/upcoming_changes/32497.improvement.rst
   Updated doc/source/glossary.rst casting definition with reference to scalar rules.
   Added a dedicated section Cast safety of Python scalars in doc/source/reference/arrays.promotion.rst documenting kind-based promotion, precision loss, out-of-bounds OverflowError, and "no" / "equiv" restrictions.

3. Unit Tests
   test_api.py: Added assertions in test_copyto_cast_safety verifying that casting="no" raises TypeError when target dtypes differ.
   test_multiarray.py: Fixed in-bounds scalar values in test_output_dtype and added test_pyscalar_out_of_bounds for np.choose.
   test_shape_base.py: Added test_pyscalar_out_of_bounds, test_pyscalar_casting_matches_copyto, and test_pyscalar_casting_matches_ufunc.
   test_stringdtype.py: Removed obsolete assertions in test_pystr_scalar_concatenate_preserves_nulls that expected "no" and "equiv" to succeed for Python strings.
   test_ufunc.py: Added casting="no" tests for scalar operands in test_cast_safety_scalar and test_resolve_dtypes_basic.

Result

1. Consistent Casting Rules:
   casting="no" and casting="equiv" now behave consistently: if converting a scalar to a non-default dtype fails under "equiv", it also fails under "no".

np.copyto(np.zeros(3, dtype="uint8"), 3, casting="no")
TypeError: cannot cast Python int to uint8 under the casting rule 'no'

2. Safe Python Scalar Handling in concatenate and choose:
   Out-of-bounds scalar values now consistently raise OverflowError rather than silently overflowing or wrapping:

np.concatenate((np.ones(2, "int8"), 300), axis=None)
OverflowError: Python integer 300 out of bounds for int8

np.choose([0], (-1, np.array([1], dtype=np.uint8)))
OverflowError: Python integer -1 out of bounds for uint8

Full Parity Across Operations:
np.concatenate(axis=None), np.choose, np.copyto, and ufuncs now share the exact same scalar casting and conversion behavior.