MLIR support for differentiating parallel loops

We've noticed some support for differentiating parallel loops was recently added (e.g., 902bf1f for `scf.parallel`), but it seems incomplete. We were wondering if there is work to more fully support differentiating parallel loops (in particular, forward mode and `scf.parallel.reduce`), either planned or on-going? Differentiating parallel loops is central to our work, which is something we could contribute to, but don't want to duplicate efforts.

@brian-kelley