dpt.sum is not doing computation in requested `dtype` on Windows

```
>>> import dpctl.tensor as dpt
>>> a = dpt.full((2, 3, 4), 123456789, dtype=dpt.int32)
>>> dpt.sum(a, dtype="f4")
usm_ndarray(-1.3320044e+09, dtype=float32)
>>> dpt.sum(dpt.astype(a, "f4"), dtype="f4")
usm_ndarray(2.962963e+09, dtype=float32)
```

Both outputs should agree. This is cause test failure observed in IntelPython/dpnp#1617