Enable control_flow_ops_py_test for Windows tests.

This was probably broken in PR #25261 which replaced np.asscalar(x[0])) with
x.item(0) for SlowAppendQIntArrayToTensorProto. This fails when x is a ndarray containing a tuple. For example:

a=np.array([42])
a=a.astype(np.dtype([('qint8', 'i1')]))
[x.item(0) for x in a]  # [(42,)]
[np.ravel(x.item())[0] for x in a]  # [42]
PiperOrigin-RevId: 238354893