Enforce dtype conservation in ufuncs that explicitly use dtype= (#7808)

* enforce requested dtypes

* add a test

* revert the handling of out=; improve test

* blacken

* make what the test should be checking more explicit

* remove unncessary initial dtype check

* fix bad kwarg

* Adjust test: simplification suggested by Julia

dask/array/core.py

@@ -4311,6 +4311,7 @@ def elemwise(op, *args, **kwargs):
 
     need_enforce_dtype = False
     if "dtype" in kwargs:
+        need_enforce_dtype = True
         dt = kwargs["dtype"]
     else:
         # We follow NumPy's rules for dtype promotion, which special cases

dask/array/tests/test_ufunc.py

@@ -474,3 +474,19 @@ def test_divmod():
     expected = divmod(arr1, arr2)
     assert_eq(result[0], expected[0])
     assert_eq(result[1], expected[1])
+
+
+@pytest.mark.parametrize("dt", ["float64", "float32", "int32", "int64"])
+def test_dtype_kwarg(dt):
+    arr1 = np.array([1, 2, 3])
+    arr2 = np.array([4, 5, 6])
+
+    darr1 = da.from_array(arr1)
+    darr2 = da.from_array(arr2)
+
+    expected = np.add(arr1, arr2, dtype=dt)
+    result = np.add(darr1, darr2, dtype=dt)
+    assert_eq(expected, result)
+
+    result = da.add(darr1, darr2, dtype=dt)
+    assert_eq(expected, result)