Fix broadcast semantics

Summary:
According to https://pytorch.org/docs/stable/notes/broadcasting.html, for two tensors, "the dimension sizes must either be equal, or one of them is 1".

When one of them is 1, broadcast happens, and the size is changed to match the other tensor.  The doc states that for each dimension size, the resulting dimension size is the max of the sizes of x and y along that dimension.

However, when the other tensor has size 0, the size of both tensors are broadcasted into 0, not 1 (max between 0 and 1). This is how numpy behaves. ATen has the same behavior as numpy.
```
>>> np.add(np.zeros([1, 0]), torch.ones([1, 1, 1]))
tensor([], size=(1, 1, 0), dtype=torch.float64)

>>> torch.zeros([1, 0]) + torch.ones([1, 1, 1])
tensor([], size=(1, 1, 0))

```

Therefore, `max` is not always true in this case. It's always changing the tensor with size 1 to match the other.

Reviewed By: manuelcandales

Differential Revision: D48167767

fbshipit-source-id: 9d65e7fad3f9fae1db184041f72a4c77e150ee39