Document how to perform benchmarks on the GPU

We should describe how benchmarks can/should be performed on the GPU (see https://trixi-framework.org/TrixiDocumentation/stable/performance/). For example being careful that all the arrays are adapted to the GPU only after the `ode` object is created, and trying to use `semi` instead of `ode.p` is incorrect, as an example.

Start with the stable performance documentation at https://trixi-framework.org/TrixiDocumentation/stable/performance/ and review its existing benchmark guidance. Add GPU-specific instructions covering when arrays should be adapted relative to creating the ode object, plus an example explaining why using semi instead of ode.p is incorrect; the documentation is done when these cases are clearly described.

Written by the indexing model from the issue text.