functionalization: fix for mutable ops with different type promotion rules (#81702)
fixes https://github.com/pytorch/pytorch/issues/81618
At some point it looks like this became broken (you can see the updated expect test looks better now, and the original was just returning a constant).
I also got a repro that was failing with an assert, that I confirmed now passes:
```
def foo(t, y):
out_1 = torch.ones(1)
return torch.add(t, y, out=out_1)
g = make_fx(functionalize(foo))(torch.tensor([1]), torch.tensor([1]))
print(g.code)
out1 = functionalize(foo)(torch.tensor([1]), torch.tensor([1]))
out2 = foo(torch.tensor([1]), torch.tensor([1]))
print(out1 == out2)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81702
Approved by: https://github.com/ezyang
