Width normalizer in cost volume fusion

[zhouxf53]

https://github.com/ardaduz/deep-video-mvs/blob/fa14288f149c5af7b2a49092f729f5c4f44517ba/dvmvs/utils.py#L62

I think it is supposed to be

width_normalizer = (width  - 1) / 2.0
height_normalizer = (height - 1) / 2.0

The reason of that is to map x from [0, width-1] to [-1, 1] to grid sampling instead of map [0, width] to [-1, 1]