Special handling for 0-D tensor

Summary:
`ET_CHECK_VALID_DIM` checks whether a tensor has a valid dimension. However, for 0-D tensors, some behaviors are different.

Accessing size/stride for dim=0 is invalid, because the valid dimension is [-0, 0) which is none.
```
torch.tensor(2).size(dim=0)
```

However, some ops allow accessing dim=0 or -1 for a 0-D tensor
```
torch.mean(torch.tensor(2, dtype=float), dim=-1)
```

Therefore, in reduce_util helper functions and ops, we need to special handle that case.

We also want to revisit this check for 0-D tensor case for ops.

Reviewed By: manuelcandales

Differential Revision: D48319644

fbshipit-source-id: 0a394ab4caccbcbb7868ccd371276e9dd047f054