torch.ops.aten invocations inconsistently cause graph break

repro

```
    def test_torch_ops_aten(self):
        # Picked an op that doesn't show up in the default list
        @torchdynamo.optimize("eager", nopython=True)
        def fn(x):
            return torch.ops.aten.absolute(x)

        fn(torch.randn(3))
```

I know how to fix this so I'm going to handle it