Future plans?

[simsurace]

I have been thinking about how to ensure KernelFunctions.jl keeps working going forward with reasonable maintenance. I would propose the following:

• Make a minor release with Julia compat bumped to 1.6 and ensure tests pass up to and including LTS (just accept that this version won't work on 1.12)
• Make a breaking release with Julia compat bumped to 1.10, moving everything AD related to package extensions. Zygote integration tests would only be run for Julia versions <1.12
• Evenutually make another breaking release where Zygote support is dropped completely

The idea being that Zygote is probably not going to be maintained anymore. Maybe the default recommendation could be Mooncake and/or Enzyme. I haven't recently tested either of them though. Last time I checked Enzyme would still fail a lot of the times, and there is a related open issue EnzymeAD/Enzyme.jl#2355. @willtebbutt what about Mooncake, is it conceivable that it would work with KernelFunctions and replace Zygote anytime soon?

[willtebbutt]

This sounds reasonable to me.

Mooncake should just work, but it might not be quite as performance as Zygote in all cases (owing to a lack of rules specific to KernelFunctions.jl). That being said, it should be quite a bit more robust, and definitely higher performance in a bunch of places where we don't have rules.