Comparing tensor subclass instance with `None` causes segmentation fault

Repro:

```python
# LoggingTensor from test_python_dispatch.py
x = LoggingTensor(torch.tensor([3.0, 4.0], requires_grad=True))
self.assertEqual(x, None)
```
 


cc @ezyang @gchanan @zou3519 @bdhirsh @jbschlosser @albanD