Remove managed memory from cuda backend (#902)

Completely remove managed memory from the gpu backend.
`managed_vector<level>` is split into:
- A device vector of structs containing meta data about each level 
- A device vector containing concatenated lengths of each branch in a level for all levels
- A device vector containing concatenated parents of each branch in a level for all levels

All other instances of managed memory in `matrix_fine` are easily replaced by device vectors  
Managed memory is also removed from the unit tests. 

Fixes #841