[Python] Some tests that are off by default use NumPy without @pytest.mark.numpy #55458

Describe the bug, including details regarding any error messages, version, and platform.

A few tests use NumPy but have no numpy mark. They belong to groups that are off by default, so the no-NumPy CI job (conda-python-3.13-no-numpy) never runs them, and the missing mark goes unnoticed. large_memory tests only run on the ARM64 macOS job, and CUDA tests only run on GPU runners; both have NumPy installed.

Affected tests (these modules set np = None when NumPy is missing):

• test_convert_builtin.py::test_auto_chunking_list_like (large_memory, uses np.ones)
• test_convert_builtin.py::test_array_from_pylist_data_overflow (large_memory, uses np.zeros)
• test_convert_builtin.py::test_array_from_pylist_offset_overflow (large_memory, slow, uses np.zeros)
• test_array.py::test_non_cpu_array (needs CUDA, uses np.arange)

With NumPy uninstalled, running these with --enable-large_memory (or on a CUDA machine) fails with AttributeError: 'NoneType' object has no attribute ... instead of skipping.

The fix is to add @pytest.mark.numpy to these four tests.

Found while reviewing #55264.

Component(s)

Python