Implement parallel band matrix solve

There are parallel algorithms for solving band matrix systems (e.g., cyclic reduction / recursive doubling). This could further speed up `ldiv!`, which mostly spends time in `band_matrix_solve!`.

Start by reading the `ldiv!` path and the `band_matrix_solve!` routine, since the issue identifies the latter as the main cost. Evaluate cyclic reduction or recursive doubling for the band systems; done means a parallel solve is implemented and `ldiv!` benefits from the improved `band_matrix_solve!` performance.

Written by the indexing model from the issue text.