Finding cells to refine may be inefficient

In #361, the line
```julia
# The "Ref(...)" is such that we can vectorize the search but not the array that is searched
elements_to_refine = searchsortedfirst.(Ref(cache.elements.cell_ids[1:nelements(dg, cache)]),
cells_to_refine)
```
was replaced by
```julia
# Find all indices of elements whose cell ids are in cells_to_refine
elements_to_refine = findall(cell_id -> cell_id in cells_to_refine, cache.elements.cell_ids)
```
While the latter is more readable and works if `cell_ids` is smaller than `cells_to_refine` (which is important for MPI, see [this comment](https://github.com/trixi-framework/Trixi.jl/pull/361#discussion_r535064054)), the former utilizes the fact that `cell_ids` is sorted.

As discussed with @sloede, this is currently not relevant for performance, but in case it eventually becomes relevant, this issue could help us identify the problem.

Start with the two cell-selection expressions shown in the issue and the discussion in pull request #361, especially the linked comment. Compare their behavior and cost when cache.elements.cell_ids is sorted and when it is smaller than cells_to_refine for MPI. Done means documenting a measured conclusion or making a justified improvement supported by relevant checks.

Written by the indexing model from the issue text.