Preserve Python dispatch keys upon copy_tensor_metadata_except_version_counter

Whether or not this is a reasonable operation to do in the presence of
subclasses is a good question in and of itself, but this fixes an
obvious invariant violation, which is that if a Tensor reports that
it is a tensor subclass, it had better have the Python dispatch key.
Previously, the dispatch key would have gotten unconditionally cleared;
now we preserve what ever the original bit was.

Signed-off-by: Edward Z. Yang <ezyangfb.com>

Pull Request resolved: https://github.com/pytorch/pytorch/pull/75644

Approved by: https://github.com/albanD

test/test_python_dispatch.py

@@ -706,6 +706,32 @@ $6 = torch._ops.aten.add_.Tensor($1, $5)''')
         x.neg()
         self.assertEqual(called, [torch.ops.aten.neg.default])
 
+    def test_set_data(self):
+        called = 0
+
+        class SubTensor(torch.Tensor):
+            __torch_function__ = torch._C._disabled_torch_function_impl
+
+            @classmethod
+            def __torch_dispatch__(cls, func, types, args=(), kwargs=None):
+                nonlocal called
+                called += 1
+                return super().__torch_dispatch__(func, types, args, kwargs)
+
+        x = SubTensor(torch.empty(2))
+        x.data
+        self.assertEqual(called, 1)
+        x.data = torch.empty(2)
+        self.assertEqual(called, 1)
+        x.data
+        self.assertEqual(called, 2)
+        self.assertIs(type(x), SubTensor)
+        x.set_(torch.empty(2))
+        self.assertEqual(called, 3)
+        x.data
+        self.assertEqual(called, 4)
+        self.assertIs(type(x), SubTensor)
+
     def test_construct_int_tensor(self):
         class SubTensor(torch.Tensor):
             pass