Fix cunumeric.arange issues

[yimoj]

Fix #531 case 1-3

Case 1 & 2:
Cause:
cuNumeric tries to create an ndarray with the shape of (ceil((stop - start)/step),) and when the value is negative we hit error from NumPy stating negative dimensions are not being allowed.

Fix:
Create empty array when the ceil is negative.

Case 3:
Cause:
cuNumeric current uses stop solely to determine out dtype and when stop is integer while the other two consist floating points there would be a precision loss.

Fix:
Infer dtype from all three arguments.

[bryevdv]

@yimoj FYI these are the two cases failing in CI (at least in the one job I checked)

   FAILED tests/integration/test_diag_indices.py::TestDiagIndicesErrors::test_negative_n[-10.5]
   FAILED tests/integration/test_diag_indices.py::TestDiagIndicesErrors::test_negative_n[-1]

[yimoj]

Doesn't seem to be a valid test, the behavior should have been xfailing if it were comparing the behavior with NumPy.

    @pytest.mark.parametrize("n", [-10.5, -1])
    def test_negative_n(self, n):
        with pytest.raises(ValueError):
            num.diag_indices(n)

>>> np.diag_indices(-10.5)
(array([], dtype=float64), array([], dtype=float64))
>>> num.diag_indices(-10.5)
(array([], dtype=int64), array([], dtype=int64))

I'll fix the test and see if there are more issues.

[yimoj]

Force push is rebase from 23.07.