The unstructured stuff changed the implementation of the parallel key
to depend on the quadrature tuple. This means that we need to make sure
that the state of the quadrature tuple is correctly all the time!

Using the same concept we use for visiting with hasattr.

This penalty accounts for the increased work in the quadrature loop.
Still a bit experimental.

This reverts commit 4f25d438.

We do this from multiple subdirs

They are only to a lesser extent important for our HPC applications
and the vectorization algorithm struggles with the complexity of the
search space.

It would be hard to implement on top of the latest feature and contradict
features projected for the near future.

By implementing repr such that instance addresses are
not part of sorting anymore

It does not yet do all the jobs it can do, but it exists and we can
run tests with it.

by introducing an order on the kernel objects and
using the first n kernels instead of an itertools.permutations
generator