Feature request: support two-sided scaling of non-batch matrices

[balos1]

The batch matrices have a scale operation which supports row scaling and column scaling (added in #1527). The scale operation for the non-batch matrices only allows scaling by a scalar. SUNDIALS would benefit quite a bit from two sided scaling support for the non-batch matrices too.

[balos1]

Another thing to note. With the normal operation mode of SUNDIALS, the modification of the matrix in-memory makes us also have to unscale the matrix after a solve. This is because we normally lag the matrix across time steps (and therefore across linear solves), but allow for the scaling vectors to change every time step. If somehow the scaling could be applied only when the matrix operator is applied (and therefore the matrix entries are not changed in-memory) that would be nice.

[MarcelKoch]

Your second point should be possible by using a gko::Composition. This is a class that takes Ginkgo operators A, B, C, ... and its apply computes A * B * C * ... * x. If you use two diagonal matrices for the scaling, then you should be able to achieve what you describe.
Please note that our preconditioners don't work out-of-the-box with the composition. You might need to explicitly generate the preconditioner (i.e. use Solver::build().with_generated_preconditioner(...)).

[pratikvn]

@balos1, can you elaborate on

if somehow the scaling could be applied only when the matrix operator is applied (and therefore the matrix entries are not changed in-memory) that would be nice.

If I understood it correctly, I guess you would like the matrix to be immutable, but the scaling to be an operation only performed within the linear solve ?