Incorrect normalization of W and H in inference for non-square input images

I found a bug in the GenWarp code when I tried inference on a non-square input image.

https://github.com/sony/genwarp/blob/main/genwarp/GenWarp.py#L396
```python
        coords = torch.stack((grid[..., 0]/H, grid[..., 1]/W), dim=-1)
```
The normalization of H and W is flipped, and should be corrected. I’ll probably make a PR sometime soon.