Fixing error in DPM solver

[Simonbiene]

I found that there is an issue with the DPM solver, when using dpmsolver++. The predicted segmentation masks have values way beyond [0,1]. As a result, clipping the mask to [0,1] typically yields a segmentation of the whole brain instead of the tumor.

The issue lies in the last diffusion step, where no corrections to the model prediction is applied. Applying "data_prediction_fn" in a slightly changed way in the last step fixes the issue.