feat: use DI for structured Jacobians

[avik-pal]

Seems to work beautifully. Haven't benchmarked yet, but the important part here is to ensure structure is preserved, since the main cost in these workloads is that Linear solving will be bottlnecked if a banded matrix is converted into a sparse matrix.

[gdalle]

Structure is indeed preserved for decompression (on the SparseMatrixColorings PR branch at least).
However, with the current implementation, a sparse matrix and its transpose will be allocated from the structured Jacobian nonetheless, as a preparation step for decompression.
This seems futile when the decompression target still has the same structure, but when it doesn't, this sparse encoding is actually useful for fast decompression. The scenario here would be someone passing sparsity_pattern::Tridiagonal but then calling jacobian! on a jac_prototype::SparseMatrixCSC.
Ultimately, coloring is a one-time cost, whereas decompression is paid at every Jacobian computation. So in SparseMatrixColorings we choose to allocate a little more during coloring, and that way we speed up more decompression scenarios down the road.

[gdalle]

Same question about cases where DI preparation is skipped

[avik-pal]

Only where the user provides an analytic jacobian. This one is a special case for scalars when DI is skipped if jacobian/jvp/vjp is provided.