Improve the example finetuning notebook

[constantinpape]

It's very unclear how to change the data loader settings to make it work for different kinds of data. In particular it's not good that is_segmentation_dataset=True is set. Also the way of val and train rois is not easy to generalize.

[constantinpape]

We should also further improve the data normalization function to deal with negative values.

[constantinpape]

Also related: we should update the training function to use train_sam_for_configuration, the default model type to be vit_b_lm and the model types in the configurations to use the _lm models.