cuda.parallel: Incorrect use of host init array in test_reduce.py

In the course of refactoring I replaced `np.asarray([0], dtype=dt)` which created 1d single element array with functionally incorrect `np.zeros([0], dtype=dt)` which creates 1d no-elements array.

`np.zeros([0], dtype=dt)` should have been `np.zeros(1, dtype=dt)`. 