Add support for including original image to segmentation decoder forward pass

[nkaenzig]

We've found in experiments that including passing the original image to the decoder can boost performance significantly, as fine-grained information about the object outlines might be lost in the embedding space rendered by vision FMs.