Implement seamless predictions

[arentkievits]

Currently, the prediction is divided into chunks if the volume does not fit into VRAM. However, this creates seams in the prediction, where mitochondria on the border of the chunk are split up. So these predictions should be somehow merged.

Note that this is less of an issue when using the CPU, but the predictions are super slow. Even then, the volume is still limited by RAM.

Todo: Find out how to do this.