Try mixing compile-time and runtime pacxx configurations
I.e.: can I compile a program with a pacxx that only includes native support, and execute it with a pacxx that includes cuda support, on a cuda GPU?
This would be really beneficial for portability