[feature request] unique to work on nd tensors via dim arg

Can we have `unique` to return unique rows and cols as well, similar to `numpy.unique`? Example usage to give an idea:
```
>>> a = torch.tensor([[1, 0, 0], [1, 0, 0], [2, 3, 3]])
>>> torch.unique(a)
tensor([0, 1, 2, 3])
>>> torch.unique(a, dim=0)
tensor([[1, 0, 0], [2, 3, 3]])
>>> torch.unique(a, dim=1)
tensor([[1, 0], [1, 0], [2, 3]])
```