Fix parallelization strategy for numerical differention

[awvwgk]

• Intel Fortran has problems to properly handle orphaned OpenMP

@ucb-mgcf, @sespic, @fabothch Do you mind rechecking your failing examples with the version uploaded at:
https://github.com/awvwgk/xtb/suites/2519893098/artifacts/54567679

Closes #458

[fabothch]

The system I investigated works now, when using GFN-FF. From my side this appears to work.

[ucb-mgcf]

The above link to the version for me to test appears to be broken. Should I test this by installing the latest master branch?

[awvwgk]

GitHub invalidates the uploaded artifacts from workflows quite quickly, the patch is included in the version on the bleed tag as well now:
https://github.com/grimme-lab/xtb/releases/tag/bleed

[ucb-mgcf]

I just confirmed that this calculation now works using the original specification, i.e. OMP_NUM_THREADS=16,1 with the bleeding-edge version. Thanks! For efficiency, would you now recommend continuing to use OMP_NUM_THREADS=1,n or go back to using OMP_NUM_THREADS=n,1?

[awvwgk]

n,1 distribution should work best, the more easy to exploit parallelism is in the displacements of the hessian instead of the individual SCC calculations. You can try to set OMP_SCHEDULE=dynamic as well, in some cases it can improve the load balancing on the threads.