Continuation of GPU acceleration in SU2: towards a GPU-native FGMRES Linear Solver

The SU2 code relies heavily on sparse linear algebra, and there is significant speed-up potential in this area through GPU-based processing. In the current implementation, the FGMRES Linear Solver GPU-offloading is being limited by frequent data copies between Host and Device. In this project, I propose to study and implement a truly GPU-native FGMRES linear solver. My idea is to refactor the code so that the largest number of operations is offloaded to the GPU for the duration of the iterative process, all while being compatible with the current level of abstraction and software logic.