Tractseg

* "Probabilistic" tracking on TOM output
* New trk format (nibabel.streamlines API). Use `--tracking_format trk` to use it.
* Option to do Mrtrix iFOD2 tracking on original FODs but filter by TractSeg masks.
* Added 3D U-Net, but not used
* Minor improvements & Bugfixes

* **Interface change**: TractSeg does not automatically flip the peaks anymore if it detects that they probably have
the wrong orientation (The peak check is only correct in 98% of the cases. In the remaining 2% it would incorrectly flip
the peaks and the user would wonder why the results of TractSeg are so bad. Therefore now the user is informed if
TractSeg thinks that flipping is needed, but he has to do it on his own and manually verify the result.) Therefore
the command line option `-deactivate_peak_check` is not needed anymore and removed.
* Bugfixes

* Reduced memory consumption of TOM (downside: `--single_output_file` for TOM not working anymore)
* Added more testing
* Number of fibers can be set as parameter
* Brain mask not needed anymore for tracking
* Make location where to store pretrained weights customizable
* Bugfixes and minor improvements

* TOM (Tract Orientation Mapping) now supports all 72 bundles instead of only 20. Downside: needs 4x more RAM (roughly 22GB).
* Code for performing [tractometry](resources/Tractometry_documentation.md)
* Added more documentation: [Best pratices for standard usecases](resources/Tutorial.md)
* Removed batchgenerators dependency. Now it might even work on windows (not tested yet!).
* **Breaking Change**: Improved interface:
* `-i` expects a peak image by default now. If you provide a Diffusion image you have to set `--raw_diffusion_input` to make
TractSeg run CSD and extract peaks
* `--output_multiple_files` is default now. If you only want one output file set `--single_output_file`
* When using a Diffusion image as input and setting `--raw_diffusion_input` the resulting peak image `peaks.nii.gz` will
not be deleted after TractSeg is finished, but stays there. Good if you want to run TractSeg again with other
output type.
* Minor improvements and bugfixes

* Super-resolution
* Uncertainty estimation
* Automatic preprocessing (rigid registration to MNI space + automatic check for correct peak orientation and flip if needed)
* Minor improvements and bugfixes
* Automatic tracking on Tract Orientation Maps
* Postprocessing and bundle specific threshold for improved results on small bundles

* Updated to pytorch 0.4
* Added bundle specific threshold
* Endings segmentations for all 72 classes
* More testing
* Bugfixes and minor improvements