Closes #31868

Problem

MaskedArray arithmetic with Python scalars produces a different dtype
than the same operation on a plain ndarray:

import numpy as np
import numpy.ma as ma

arr = np.array([100], dtype=np.int8)
masked = ma.masked_array([100], dtype=np.int8)

print((arr + 50).dtype)     # int8  (Expected, respects NEP 50)
print((masked + 50).dtype)  # int64 (Bug)

This was due to:
_MaskedBinaryOperation.__call__ calls getdata() on both operands before calling the ufunc. For Python scalars, getdata() has no
._data to grab, so it falls back to np.array(50) which creates a 0-d int64 array. Under NEP 50 that's a "strong" type, so the ufunc
promotes int8 to int64. The standalone power() function has the same issue.

Fix

Skip getdata() for Python scalars (int, float, complex) in the two call sites, passing them straight to the ufunc where NEP 50
treats them as weak types.

Test & Enviornment

Added test_scalar_arithmetic_preserves_dtype_nep50 covering int8,float32, four operators, reversed operand order, and numpy-scalar promotion.

• OS: Darwin 25.5.0 (arm64)
• Python: 3.14.0
• NumPy: 2.6.0.dev0+git20260706.7e381e7

$ spin test numpy/ma/tests/test_core.py -- -x -q
...
=========================== short test summary info ===========================
SKIPPED [1] numpy/testing/_private/utils.py:2692: Could not determine available memory...
XFAIL numpy/ma/tests/test_core.py::TestMaskedConstant::test_coercion_unicode - See gh-9750
XFAIL numpy/ma/tests/test_core.py::TestMaskedConstant::test_coercion_bytes - See gh-9750
4170 passed, 1 skipped, 2 xfailed in 2.37s

AI Disclosure

AI was used only to strictly format this PR for maintainer readability and ease I also Used an AI assistant to help trace the call chain from MaskedArray.__add__ through _MaskedBinaryOperation.__call__ to
getdata, and to confirm the root cause.

The diagnosis, fix, and tests were written and verified manually.