Avoid reallocation for self residual norm calculation

[yhmtsai]

In most of our solver, we need to calculate the actual residual norm by b - Ax when required because these solvers only have the implicit residual norm, which is updated from the previous iteration such that it might not be correct enough for residual norm like CG.

This PR avoids reallocation the internal rhs in each iteration.
Note. we still have the allocation when generating the criterion, but it also requires storing the criterion in the solver and it is more critical when we setup some residual-norm aware smoother not the usual setup.

[MarcelKoch]

Isn't this the same as

Suggested change

	using dense_b_type = std::remove_cv_t<
	std::remove_reference_t<decltype(*dense_b)>>;
	using dense_b_type = std::decay_t<decltype(*dense_b)>>;

[MarcelKoch]

Either these blocks need a SCOPED_TRACE or the above is extracted into a new test fixture and the blocks are fully separated.

[upsj]

The aim looks good, but I would like to aim towards reducing mutable usage in our classes. That would also make thread-safety #996 easier. If we can collect the thread-unsafe parts of Ginkgo in a small number of places, like the dense caches and workspaces, then we only need to make sure those work safely across multiple threads, and the rest follows automatically. Without mutable, we can also rely on const-ness implying thread-safety.

[upsj]

I think we should avoid adding more mutable functionality, and instead either find a way to include the stopping criteria in the workspace setup, or hide it behind an opaque pointer, so at least there is no mutability in the public interface. We could do this by adding a workspace parameter to the check function and having the old version create a temporary workspace.