Squashed merge for fine matrix solver (#640)

Add a new Hines matrix solver implementation for the GPU that can solve a single tree in parallel with multiple threads. It replaces the interleaved solver, which used a single thread to solve each matrix.
Branches with the same common root in the tree can be solved independently on each of the forward and backward solution passes. 

* Add a matrix storage type, `arb::gpu::matrix_state_fine` that stores the branches of multiple trees for efficient backward and forward substitution.
* Extend the `arb::tree` data structure to support operations for choosing a new root node and determining a root node which minimises the maximum distance between the root and any of the trees leaves. 
* Implement code for rebalancing a set of matrix trees, a.k.a. a "forest" of trees.
* Add CUDA kernels for efficiently performing matrix assembly and matrix solution steps.
* Add CMake option `ARB_WITH_GPU_FINE_MATRIX` for toggling the new solver (default `on`).