* Semantic-segmentation loss calculation: fix buffer usage on multi-GPU training

* Review fix: make the work buffer live longer