Generalize AssmbScatr, ConjGrad, DiagPrecon, DirBndCond, FwdTrans, HelmSolve, Matrix, NeuBndCond operators to allow multiple implementations.
Remark: Previous version of those operators committed to specific implementation. However, for CUDA, repeated copies between device and host should be avoided. Consequently, each of the operators mentioned above should probably have their own CUDA implementation.
Note: StdMat may not be the relevant name for some operators but API should at least allow for a CPU and a GPU implementations.