ENH: add support for np.nanargmax and np.nanargmin - #10808

Part of #4074. Revives #8688 (original implementation by @Tobi995, closed
stale), reworked following the refactor suggested by @guilhermeleobas in
that review: a single shared implementation factory plus one combined
test, instead of two near-duplicated functions and tests.

Problem

np.nanargmax / np.nanargmin have never been supported in nopython
mode; jitting them fails at typing time:

numba.core.errors.TypingError: Failed in nopython mode pipeline
(step: nopython frontend)
Use of unsupported NumPy function 'numpy.nanargmax' or unsupported use
of the function.

What changed

• Adds @overload(np.nanargmax) and @overload(np.nanargmin) to
  numba/np/arraymath.py. Both share a small factory,
  build_array_nanarg_impl(name, operator), that returns a
  register_jitable implementation doing a single NaN-skipping pass over
  the flattened array and raising NumPy's exact error messages for empty
  and all-NaN input.

• Integer and boolean dtypes simply delegate to the existing
  np.argmax/np.argmin implementations, since NaN is not representable
  there and NumPy treats such input like the plain functions. The axis
  argument is handled by reusing the same helper those functions use,
  build_argmax_or_argmin_with_axis_impl. The
  numpy_version < (1, 22) guard from the original sketch is no longer
  needed as main already requires 1.22.

• Tests go into numba/tests/test_array_reductions.py as one combined
  parametrised test covering float and integer dtypes, axis=None plus
  positive/negative axes on 2-D and 4-D input, and the ValueError paths.
  Complex and datetime64/timedelta64 dtypes are still unsupported, happy
  to follow up with those if wanted. Also adds a release note under
  docs/upcoming_changes/.

Verification

TDD, red first (against the untouched tree):

$ ./runtests.py numba.tests.test_array_reductions.TestArrayReductions.test_nanargmax_nanargmin -m 1
FAILED (errors=1)
# numba.core.errors.TypingError: Use of unsupported NumPy function
# 'numpy.nanargmax' or unsupported use of the function.

Green after the implementation:

$ ./runtests.py numba.tests.test_array_reductions.TestArrayReductions.test_nanargmax_nanargmin -m 1
Ran 1 test in 30.073s
OK

Full touched-file suite:

$ ./runtests.py numba.tests.test_array_reductions -m 8
Ran 448 tests in 93.130s
OK (skipped=1)

flake8 numba/np/arraymath.py numba/tests/test_array_reductions.py: clean.

Environment: Python 3.12.8, numpy 2.5.2, llvmlite 0.50.0.dev2 (numba dev
index), numba main @ 3598bdd.

Notes for the reviewer

• The flatten impl keeps a single pass with a sentinel (max_idx = -1);
  the first non-NaN element is taken unconditionally via the max_idx < 0
  short-circuit.
• Ties keep the first occurrence, matching np.argmax/np.argmin
  semantics.