filter out death test on daint single_node_release

- build-catalogue
  - now installed by pip
  - enable GPU-backends
  - isolate from being installed in a non-default dir
  - rename to arbor-build-catalogue
  - allow hand-written C++ mechanisms to be linked in the same catalogue.
- CMake
  - enable PIC on arbor always to allow linking to shared objects
  - use relative paths when configuring a-b-c, to make relocation less of a problem
- Python
    - ensure a-b-c is installed by pip, along with headers and libarbor.a
    - throw errors if tools (cmake scripts, arbor package, modcc) missing
    - transition setup.py setuptools -> skbuild.
- Wheels
    - Add NML2 support to wheels
    - scripts/build-wheels.sh builds wheels, in principle valid for submission to PyPI, on your own hardware. This should be kept in sync with .github/workflows/ciwheel.yml
    - scripts/patchwheel.py corrects the rpath in the libraries in the wheels, working around a bad interplay between auditwheel and skbuild, see pypa/auditwh...

- Bump CMake to 3.18 for better CUDA support.
- Clean-up our use of said support.
- Remove old cruft and workarounds.
- Maintenance in CI: enable MacOS 11; bump CMake versions to 3.18 / 3.22; bump bors CMake to 3.18

Fixes #1790 

Reduce number of parallel build jobs to reduce risk of OOM in CI builds.

Add a CI test for ARM-SVE

- compiles and runs the unit tests with vectorization enabled for the `armv8.2-a+sv` architecture
- uses a dedicated gitlab runner on ault.cscs.ch

Fixes #1498

In response to:

https://gizmodo.com/u-s-federal-investigators-are-reportedly-looking-into-1846707144

Rationale:
* CodeCov didn't detect the breach for two months
* They took a further 2 weeks to alert users after that
* Their efforts to address the issue after that have been painful to watch
  * https://github.com/codecov/codecov-action/issues/281#issuecomment-823407167

- remove repeated CMAKE_BUILD_TYPE
- bump lcov to 1.15
- use new --exclude flag to skip generated sources

With gcc 8 as a compiler + a new version of lcov (#1350) codecov started tracking generated source files and coverage dropped to < 60%. Not sure if this is a bug in lcov or not. I've added these new flags for lcov: `--exclude "*/ext/*" --exclude "*/test/*" --exclude "*/mechanisms/*"` to fix it.

* Make Gitlab CI push images to CSCS' JFrog registry reg.giuv.cscs.ch so that sarus can pull images on Daint again. 
  This URL is probably temporary until JFrog officially goes into production later this month.
  Unfortunately JFrog is currently behind the firewall.
* Upgrade g++ to g++8 on the Docker image.
* Only require libxml when compiling with NeuroML support.

Update all Python scripts used during build/testing/running Arbor to explicitly use Python 3.

This is copied from DLA Future. It should set the CI status on the commit from the PR instead of just on the Bors commit. This way the gitlab CI status also appears in the checks section below the PR.

This is a bit experimental. It builds arbor with `-fprofile-arcs -ftest-coverage` and creates a docker image with just the project binaries, lcov, gcov and generated `.gcno` files. This should still be reasonably small since there are no object files etc. When run on daint, `.gcna` files are produced (should be thread / process safe according to gcc's manual) inside of the container, and `lcov` is used to combine all those things into a single file with a random name in the mounted git project folder (hopefully this solves clashes with multiple nodes). After all tests are run, the combined reports are uploaded to codecov.io.

Combine badges on readme and move ci git submodule to `ext/`

* Add Dockerfiles for MPI + CUDA test builds on Daint
* Add bors.toml for CI on Daint
* Add the separate ci repo as submodule for convenience
* Add a CI badge for master to the top of the readme