A better memory layout for IMEX methods

At the moment, we store the variables as follows `du[var, i, j, k, element]`, where `i`, `j` and `k` loops over the local node of the given `element`. This structure makes quite hard to implement IMEX schemes (implicit in the vertical direction), because accessing to the vertical direction components is not straightforward and quite complex. Some prototype of IMEX schemes in TrixiAtmo (not yet public) at the moment relies on defining a view of the current `du` structure, through a mapping that transforms the structure as `du_wrap[var, i, element_z, element_horizontal]`, where `i` is the local node in the vertical direction, `element_z` are the element in the vertical direction, and `element_horizontal` are the elements in orthogonal direction to the vertical one (that includes $$x$$ and $$y$$ for a 3D example). (Similarly to what CGDycore.jl does) On the other hand, changing the memory layout of `du` could reduce the performance of the `Trixi.rhs!` (I'm not actually sure about that). Therefore, that may also not be the best option or solution.