Add complex autograd support for torch.cross

```
>>> a = torch.randn(4, 3, dtype=torch.cdouble, requires_grad=True)
>>> b = torch.randn(4, 3, dtype=torch.cdouble, requires_grad=True)
>>> torch.cross(a, b, dim=1)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
RuntimeError: cross does not support automatic differentiation for outputs with complex dtype.
```

cc @ezyang @anjali411 @dylanbespalko @mruberry @aocsa