The parenthesis in the macro UPDATE and INTERPOLATE
are causing problems when compiling on windows.
See arcadelab#26
for more details.

The computation of lps_from_ijk turned out to be harder
than anticipated. Reason for that is the immensely complex
and horribly covered dicom standard. Trying to use nibabel
next to outsource all the error handling to another lib.

Using nibabel.nicom.dicomwrappers.MultiframeWrapper to
construct the affine transform from the available dicom
tags.

Evidently this is complex enough to _not_ maintain it
in this project. See this post to get a feeling for the
complexity: https://nipy.org/nibabel/dicom/dicom_orientation.html#defining-the-dicom-orientation.
The used class warns, that the exposed functionality
is not well tested. I will verify that it works with
our datasets.

Using nibabel.nicom.dicomwrappers.MultiframeWrapper seems
not as well fitting as expected. The class was conceptualized
for phillips dicoms and there seems to be a non-trivial
difference.

Decided to implement the affine transform computatio from scratch
following these resources:

https://nipy.org/nibabel/dicom/dicom_orientation.html#defining-the-dicom-orientation
https://mrohleder.medium.com/coordinate-systems-in-medical-data-science-a-rant-90394f60b27
https://github.com/nipy/nibabel/blob/master/nibabel/nicom/dicomwrappers.py

Same number of parameters as from_nifti() now.

The difference in indexing between dicom and this
framework were previously dealt with in the affine
matrix by switching the columns of i and j.

Now, we change the internal layout of the supplied
volume prior after loading and leave the affine
transform untouched.

Grouped the dicom tags by per-frame and per-slice attributes for
better readability.

Removed not needed InvalidDicomError import.

The parentheses in the CUDA kernel are a different topic and should
be dealt with in a seperate thread. Removing that commit from the
history of this branch.

This reverts commit 8f45483.