"invalid context handle" crash on Cuda >= 11

End-to-end reconstruction crashes on machines with Cuda 11 (not only power9 machines) with a Cuda "invalid context handle" error.
This happens when a new pipeline instance is created and the "old" pipeline instance is destroyed.

The nvidia-persistenced daemon has been restarted with the same configuration as on the SLURM cluster. The error is likely due to scisoft15 configuration, although it would be good to find the cause.

The cause is probably the cuda context management by CudaProcessing classes.