Tractseg

* Update documentation
* Update 3d tract visualization in `plot_tractometry_results` to use fury
* Updated pytorch to 1.8.1

* 3D plot of streamlines with coloring according to tractometry FA
* Add pretrained weights for XTRACT tract definitions
* Set range of y-axis in `plot_tractometry_results` (thanks to elder-mama)
* Make `--preprocess` also work for endings_segmentation and TOM
* Simplified dockerfile
* Minor improvements

* Available on pypi
* Manually specify which bundles to track
* Minor improvements

* **Interface change**: The option `--bundle_specific_threshold` was removed. TractSeg checks itself now if CA or FX
are incomplete and then applies a lower threshold.
* **Interface change**: Postprocessing is activated by default now. If you want to deactivate is use `--no_postprocess`.
* minor improvements & Bugfixes
* FP16 training (increased training speed)
* Tractometry more testing and bugfix
* Tractometry now uses a far more advanced option to sample e.g. the FA along the tracts.
* Statistical analysis for tractometry data
* Python 2 not actively supported anymore (because dipy 1.0.0 does not support python 2 anymore)
* Add rotation (now peaks are also properly rotated) to data augmentation.
* Applies signs of affine to data if array not oriented like MNI data (needed to properly work with `fslreorient2std`)
* '--preprocess' will move output back to subject space.
* Updated weights for tract segmentation, endings segmentation and density regression (slightly increased
performance; now also trained with rotation during data augmentation; TOM not trained with rotation yet)
* Pretrained model also works with bedpostX peaks (instead of CSD peaks) (segmentation accuracy is the same)

* Increase training speed roughly by factor of 2 by using pin_memory and non_blocking for pytorch and by
cropping all non-brain area from the input images (requires preprocessing of the training data using
`tractseg/data/preprocessing.py`).
* Works with newer version of batchgenerators (Note: DataAugmentation slightly changed)
* Support bedpostX input
* Support aPTX tract definitions (but no pretrained model yet)
* Refactor `--preview`. Works without vtk now.
* Add plateau LR schedule to training
* Add API for mrtrix FACT tracking on TOMs
* Fix bug in rotation of bvecs when using option `--preprocess`.
* minor improvements
* Update TOM model and pretrained weights (Only angle in loss instead of angle and length. Gives slightly better
peak orientations.). Improved peak orientations allows for slightly less sensitive probabilistic tracking: lowering
stddev from 0.2 to 0.15.

* Tracking on best original peaks or on weighted mean of best original peaks and TOMs (non-public interface).
* **Interface change**: All tracking related commands (whenever you used `--track`) are not part of `TractSeg` anymore
but now are combined under `Tracking`. Moreover the option `--filter_tracking_by_endpoints` is now activated per
default. If you want to deactivate is use `--no_filtering_by_endpoints`.
So the following command

TractSeg -i peaks.nii.gz --output_type TOM --track --filter_tracking_by_endpoints

becomes

TractSeg -i peaks.nii.gz --output_type TOM
Tracking -i peaks.nii.gz

* Works with pytorch 1.0 now
* Bugfixes and minor improvements