CUDA solver support

[ritukeshbharali]

It would be nice to add some GPU solver interfaces. Recently, NVIDIA has released their first generation GPU-accelerated Direct Sparse Solver, cuDSS. AMGx has been around and a solver interface has been written already, but not thoroughly tested. Here is a list of things to be done:

• Wrapper to cuDSS
• Test cuDSS on the FE problems (sym, unsym)
• Write tests for cuDSS, AMGX (small problems, that checks whether they are working)
• New makefile

[ritukeshbharali]

NVIDIA has not yet released a stable cuDSS version (0.2.1), and most solver configuration options are absent. Hence, such configuration options are missing in the wrapper too. cuDSS works well for solid mechanics problems (e.g., linear elasticity, phase-field) but not for the saddle point problems in poromechanics.

AmgX by default is a distributed memory solver, however, the wrapper operates only on shared-memory space. This is because Jive has no notion of global dof numbering, see #18. AmgX also works well for problems that are not of saddle point nature.