I have a couple of question and suggestions about things I noticed during moving the initialization.

1. SU2_DOT did not call omp_initialize(), it does now after moving omp_initialize() to the driver constructor. Does it actually run any hybrid AD stuff?
2. The number of OpenMP threads can be specified for SU2_CFD via the -t parameter, but for SU2_DOT and the python wrapper I saw no way to configure it. Making the number of threads a parameter of the driver constructors and calling omp_set_num_threads in the constructor of CDriverBase seems like a natural way to enable this for all flavours of SU2. Would that be a nice addition?
3. I couldn't find any python wrapper tests among the parallel AD and hybrid AD tests. Could any of the serial AD python wrapper tests serve as a basis for a parallel AD or hybrid AD one?

Thanks for your approval @pcarruscag. Could I have your opinion on the questions? E.g., I could add 2. as part of this PR, now that I already looked into it.

Ah sorry,
SU2_DOT mostly only uses OpenMP in the passive linear system, I'm not sure if @thomasdick Sobolev smoother has some recorded part that involves OpenMP.
For the python wrapper I usually set the number of threads with env var OMP_NUM_THREADS It may be a bit confusing for most users having to specify the number of threads when instantiating the driver 🤔
I think disc_adj_flow would be the easiest to adapt. Those tests print a sensitivity at a particular vertex of a marker to be checked by the regression script, so you would have to make sure that's only printed on one rank (for parallel AD).
disc_adj_fea should also be fine for hybrid AD.

Thanks for your feedback! I agree, OMP_NUM_THREADS is better than having the number of threads in the constructor. I adapted disc_adj_flow and disc_adj_fea for MPI and added both to parallel AD and hybrid AD tests.