Only require numpy for the vectorized UDF path

[yvaishalirao]

Fixes duckdb/duckdb#17052 (duckdb/duckdb#17052)

create_function() was unnecessarily requiring numpy even for native UDFs that never use it. GetFunction() in src/python_udf.cpp imports/version-checks/warms-up numpy before determining if the vectorized mode was chosen. Only the vectorized path actually needs numpy, since it round-trips through pyarrow — the native path never touches either.

Fix is one line, scoping that block inside if (vectorized).

Two new tests cover both directions:

• test_scalar_native.py::test_native_udf_without_numpy
• test_scalar_arrow.py::test_arrow_udf_requires_numpy

Both run in a subprocess rather than monkeypatching sys.modules in-process — numpy's import cache (PythonImportCacheItem) lives for the whole interpreter process, so if numpy's already been imported by another test in the same worker, an in-process monkeypatch won't make the cache forget it. A fresh subprocess sidesteps that.

Verified the test actually catches the regression by stashing the fix and running it both ways, numpy uninstalled in both:

Build	numpy installed?	Result
Pre-fix	No	fails — 'numpy' is required for this operation, but it wasn't installed
Post-fix	No	passes