Inconsistent results from torch.triu on GPU compared to CPU

On recent PyTorch master (built from source; 0.4.0a0+4970e73) on a P100, CUDA 9:

```
>>> torch.triu(torch.FloatTensor([1]).expand(3, 3))

 1  1  1
 0  1  1
 0  0  1
[torch.FloatTensor of size 3x3]

>>> torch.triu(torch.cuda.FloatTensor([1]).expand(3, 3))

 1  1  1
 1  1  1
 1  1  1
[torch.cuda.FloatTensor of size 3x3 (GPU 0)]
```