Revert "faketensor: prevent deepcopy from cloning FakeTensorMode (#104476)"

This reverts commit c54afea6ee.

Reverted https://github.com/pytorch/pytorch/pull/104476 on behalf of https://github.com/jeanschmidt due to sadly it is breaking internal tests, and I can't coordinate a FF due to timezone differences ([comment](https://github.com/pytorch/pytorch/pull/104476#issuecomment-1661808343))

test/test_fake_tensor.py

@@ -1014,15 +1014,6 @@ class FakeTensorOperatorInvariants(TestCase):
 
             self.assertEqual(ref.size(), meta_out.size())
 
-    def test_module_deepcopy(self):
-        import copy
-        from torch._guards import detect_fake_mode
-        with FakeTensorMode() as m:
-            lin1 = torch.nn.Linear(2, 2)
-            lin2 = copy.deepcopy(lin1)
-            all_params = list(lin1.parameters()) + list(lin2.parameters())
-            curr_mode = detect_fake_mode(all_params)
-            self.assertTrue(curr_mode is m)
 
     @skipIfRocm
     @unittest.skipIf(not RUN_CUDA, "requires cuda")

torch/_subclasses/fake_tensor.py

@@ -1028,16 +1028,6 @@ class FakeTensor(torch.Tensor):
     def from_tensor(t, fake_mode):
         return fake_mode.from_tensor(t)
 
-    # FakeTensorMode is meant to be a singleton, so deepcopying
-    # should not introduce a fresh mode.
-    # This just implements the "default" deepcopy, but without deepcopying
-    # the fake_mode.
-    def __deepcopy__(self, memo):
-        result = torch.Tensor.__deepcopy__(self, memo)
-        assert isinstance(result, FakeTensor)
-        result.fake_mode = self.fake_mode
-        return result
-
     @classmethod
     @count
     def __torch_dispatch__(cls, func, types, args=(), kwargs=None):