GPU support

We should explore porting MESA to GPUs. Here is one approach for offloading some of our most intensive calculations: the block tridiagonal matrix solve

cuSparse has a tridiagonal solver we could use:
https://docs.nvidia.com/cuda/cusparse/index.html?highlight=tridiagonal#tridiagonal-solve

For interfacing Fortran with GPUs, we can use CUDA Fortran to call directly. Here’s the documentation for doing that with cuSPARSE:
https://docs.nvidia.com/cuda/cusparse/#cusparse-fortran-bindings

Start by locating MESA's block tridiagonal matrix solve, then review the linked cuSPARSE tridiagonal-solver and CUDA Fortran documentation. The issue does not name files, tests, or a concrete completion target; scope and validation for the GPU port would need to be established first.

Written by the indexing model from the issue text.