GitLab

- Update CI stage to reflect update
- Update ci_scripts to reflect update
- Reformate code with clang-format-10
- Update contribution guide

The width and height of a detector in the CircleTrajectorGenerator is based on the size of the volume. To avoid
undersampling of the corners of the volume, the size of the detector is now the size of the volume
scaled by sqrt(2). This is closer to how matlab behaves

restructured Docerfiles

- Bump CMake to 3.18.2
- Bump oneDNN to 1.6.2
- Bump to Ubuntu 20.04
- Seperate images for clang-format, clang-tidy, coverage, Cuda build
  with clang, gcc and clang with python bindings
- Dockerfiles for clang and gcc rely on the package manager for the
  compilers
- Updated test scripts, they test more and should error our earlier
- Nightly docker job created
- Docker build script updated
- Include CI job, to build docker images on changes

Coverage includes build folders of python bindings, which pull down the coverage unnecessarly, therefore they are not generated for the coverage CI job

Turn python bindings off for checking with address sanitizer, as they create problems either with linking, or many false positives (see #83 for more info)

- Pipeline includes jobs for installing elsa, cloning elsa_testing and building it
- Adjustments made to install process, such that elsa_testing can be build and everything is working

- Chaching is introduced for for all jobs, it caches for a branch.
only submodules are chached, as build folders should be cached using
artifacts, which is also done. We also cache the .git/modules folder
as anyway we have to reclone everything everytime.
- CMake UpdateGitSubmodules now goes over all submodules and check
individually if they have to be initialized, or just updated.
Also work around a specific error in the CI pipeline, if
.git/modules/thirdparty/<submodule> is not present (which by default
dosent get cached), we have to deinit it and then reinit it.

Add benchmark folder to clang-format check; Add target build-examples; Add targets for building benchmarks and examples to CI pipeline

Versions of clang>=9 evaluate constexpr structs as syntactically
incorrect initializer lists. Force conversion to overloads on the Python
side for all default arguments that are structs.

The natural  numpy equivalent of the DataContainer is the Fortran array.
Supporting arbitrarily strided buffers can be tricky. Require np.array
as input and force conversion to Fortran order if a different type of
array is supplied. This removes the indexing convention discrepancy
between numpy and elsa.