Adding reverse mode AD support

Larger PDE solves may benefit from a reverse mode AD and also helps with adding support for neural networks as surrogates or part of training the PDEs themselves, therefore I was wondering what it would take to add Zygote support here. It also comes with a forward mode which may be something that can be used as well.

The issue proposes adding Zygote support for reverse-mode AD to Trixi.jl, with forward mode also mentioned for larger PDE solves and neural-network use cases. Start by locating the PDE solve entry points and assessing how Zygote would interact with them; done should mean a defined, working AD path for the stated use cases.

Written by the indexing model from the issue text.