GPU compatible `ForestBWG`

having the tree management itself on the gpu is quite young and limited. So this would be nice as a collaborative research project. The forest/tree traversal algorithms need to change for this.

some inspiration from non fem code:
- https://www.sciencedirect.com/science/article/pii/S0010465525000463
- https://github.com/lanl/CLAMR

forest on cpu and compute on gpu is already quite established (and should already be possible with the known restrictions based on @termi-official GPU branch)