MAINT: ban setting ndarray.dtype (numpy#47)

See discussion in https://github.com/numpy/numpy-stubs/issues/7.

Though this is valid NumPy, users should be using `ndarray.view`
instead of setting the dtype, and setting the dtype prevents sensibly
making `ndarray` generic over dtype because of situations like this:

```
x = np.ones((2,), dtype=np.int64)  # type is ndarray[np.int64]
x.dtype = np.bool_  # What is the type now?
```

If someone really wants to do this, they will now have add an ignore,
which should make it clear that type safety is going out the window.

Author: person142

.gitignore

@@ -3,4 +3,5 @@
 __pycache__
 numpy_stubs.egg-info/
 venv
-.idea
+.idea
+**~

numpy-stubs/__init__.pyi

@@ -297,8 +297,6 @@ class ndarray(_ArrayOrScalarCommon, Iterable, Sized, Container):
     ) -> ndarray: ...
     @property
     def dtype(self) -> _Dtype: ...
-    @dtype.setter
-    def dtype(self, value: _DtypeLike): ...
     @property
     def ctypes(self) -> _ctypes: ...
     @property

tests/fail/ndarray.py

@@ -0,0 +1,11 @@
+import numpy as np
+
+# Ban setting dtype since mutating the type of the array in place
+# makes having ndarray be generic over dtype impossible. Generally
+# users should use `ndarray.view` in this situation anyway. See
+#
+# https://github.com/numpy/numpy-stubs/issues/7
+#
+# for more context.
+float_array = np.array([1.0])
+float_array.dtype = np.bool_  # E: Property "dtype" defined in "ndarray" is read-only