Totalsegmentator

* fix critical bug in `body` task postprocessing: sometimes all foreground removed

* allow more than 10 classes in `--roi_subset`
* bugfix in `appendicular_bones` auxiliary mapping
* in multilable output only show classes selected in `--roi_subset` if selected
* make statistics work with dicom input

* add option `--v1_order` to use the old class order from v1

* 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

* download all weights from github releases instead of zenodo

* remove verbose print outs not needed
* add helper script for manual setup
* add fast statistics
* download weights from different server for faster and more stable download
* fix `requests` version to avoid `urllib3` openssl error
* minor bugfixes