* train models with nnU-Net v2 (nnunet_cust dependency no longer needed)
* roi_subset a lot faster, because cropping with 6mm low res model to roi first
* more classes and improved training dataset (for details see `resources/improvements_in_v2.md`)
* bugfix to make cli available on windows
* bugfixes in dicom io
* add `--skip_saving` argument
* automatic tests on windows, linux and mac
* statistics are not calculated anymore for ROIs which are cut off by the top or bottom of the image (use `stats_include_incomplete` to change this behaviour)
* add postprocessing for body segmentation: remove small blobs
* use dicom2nifti for dicom conversion instead of dcm2niix because easier to use across platforms