Floating-point discrepancy between Ubuntu and MacOS tests for `DGMulti` prism solver

In #2259, it was noted that for `Wedge` elements with a scalar polynomial degree (i.e. using the elixir https://github.com/trixi-framework/Trixi.jl/blob/dgmulti-mesh-io-support/examples/dgmulti_3d/elixir_advection_tensor_wedge.jl with `polydeg = 3` instead of the default `polydeg = (3, 4)`), Ubuntu and MacOS give `linf` error values which differ a bit more than usually observed for double-precision arithmetic. Relaxing `atol` to 1e-10 gets the tests to pass, but it is still unclear what the cause of such discrepancies is; decreasing the time step helps only marginally (even with a drastically smaller CFL and/or switching to a fixed step size, the tests do not pass under the default tolerances for `linf`).

Start with examples/dgmulti_3d/elixir_advection_tensor_wedge.jl, changing polydeg to 3 as described, and run the Ubuntu and macOS tests with their default tolerances. Compare the linf errors while checking the DGMulti prism solver and time-step settings; done means the discrepancy's cause is identified and the tests pass consistently or the tolerance change is justified.

Written by the indexing model from the issue text.