Totalsegmentator

* major update to all MR models: double the number of training subjects
* add `tissue_4_types` task: add intermuscular fat class
* add `vertebrae_mr` task: numbered single vertebrae segmentation in MR images (for CT this is already part of the `total` task)
* add `appendicular_bones_mr` task: add appendicular bones segmentation for MR images
* add `thigh_shoulder_muscles_mr` task: add thigh and shoulder muscles segmentation for CT and MR images
* add `vertebrae_body` with new class `intervertebral_discs`
* add `body_mr` task: add body segmentation for MR images
* add `kidney_cysts` task: greatly improved kidney cyst segmentation compared to the `kidney_cyst` class which is part of `total` task
* add `breasts` task: add breast segmentation
* add `oculomotor_muscles` task: add oculomotor muscles model
* add `lung_nodules` task (thanks to [BLUEMIND AI](https://bluemind.co/))
* update `coronary_arteries` task: increased number of training subjects, including non-contrast images
* add option to remove small connected components in postprocessing
* add `totalseg_get_modality`: estimate modality (CT or MR) from input image
* removed `rt_utils` and `p_tqdm` dependency
* change pi_time threshold for arterial late phase from 50s to 60s

* add brain structures
* add liver vessels
* greatly improved phase classification model

* Bugfixes
* add headneck structures

* also return statistics from python api
* add `totalseg_get_phase`
* major bugfix: rib labels were in wrong order
* hide nnunetv2 2.3.1 warning: `Detected old nnU-Net plans format. Attempting to reconstruct network architecture...`
* add mr models

* Bugfix: add flush to DummyFile
* Require python >= 3.9 in setup.py
* properly add `vertebrae_body` model
* add `--roi_subset_robust` argument
* add `--fastest` argument
* allow `mps` as device (but not supported by pytorch yet)
* add inline python version requirement for `requests` package
* if input spacing same as resampling spacing then skip resampling
* from python api also return nifti with label map in header
* input to python api can be a Nifti1Image object or a file path
* upgrade to `nnunetv2>=2.2.1`
* for `total` task use nnU-Net `step_size=0.8` instead of `0.5` for faster runtime while only decreasing dice by 0.001
* minor edits and bugfixes

* downgrade nnunet to 2.1 to fix bug in `fast` model