Args for ODs methods with multiple backends

[mauicv]

I feel like we need to revisit how we spec backend-specific arguments. E.g., one needs to carefully read the docstrings (and assume they are correct!) to see which arguments are relevant for the backend choice. This list would become ever more confusing with more backends, so perhaps we need to consider a hierarchical approach, e.g. **pytorch_opts, **sklearn_opts.

Originally posted by @jklaise in #736 (comment)