Optimizing gemm usage in splines

[blegouix]

In SplinesLinearProblem2x2Blocks::solve(), two gemm operations are performed, involving bottom-left and top-right corners. Those are stored as dense matrix (2D Kokkos::View). However, the bottom-left corner contains a very small number of non-zeros (here 6 for a 4x96 block):

while the top-right corner has the following structure (once factorized as Q^-1*gamma):

Of course, dense storages for those blocks and dense gemm usage are suboptimal. We could consider improving this using conversion to sparse format (removing zeros) during factorization and using some spmv.

@yasahi-hpc

[blegouix]

I think this: https://github.com/kokkos/kokkos-kernels/blob/develop/sparse/src/KokkosSparse_spmv_team.hpp

Should suit the needs once we have the Csr