multiply_dimensionwise: mortars etc.

Once https://github.com/JuliaLang/julia/pull/37444 is merged into mainstream Julia, we should revisit the choice of `MortarMatrix` in DG methods. Currently, we use `mortar_forward_upper::SMatrix` etc. which basically disables LoopVectorization in `@tullio` there. Once the performance of getting pointers to `view`s is increased, we should check whether it's better to use `mortar_forward_upper::Matrix` with LoopVectorization instead.

Xref https://discourse.julialang.org/t/tullio-loopvectorization-pointer-becomes-slow-for-views-because-of-base-memory-offset-for-ndims-5/46178