**BREAKING**: Move AD to package extensions and drop Zygote tests for Julia 1.12 and up

[simsurace]

Summary
We move AD-related code to package extensions for cleaner separation of essential vs. nonessential parts of the package.
Zygote is broken on Julia 1.12, so we mark these tests as broken to have less CI noise and not miss actual breakage.
This was discussed in #585.

What alternatives have you considered?
We could have left the rules in the main package and simply marked the tests as broken. But since the AD code is not needed when one does not load an AD package that benefits from it, this is arguably cleaner.

Breaking changes
If there ever was a promise that KernelFunctions would work with Zygote, it is now broken on Julia 1.12. Zygote is still expected to work and tested on Julia LTS. The tutorial example mentions this, but right now we don't test with any other AD packages. This can be done in a future version.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.46%. Comparing base (8c6ac8c) to head (f1bec84).
⚠️ Report is 1 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #592   +/-   ##
=======================================
  Coverage   94.46%   94.46%           
=======================================
  Files          52       52           
  Lines        1463     1463           
=======================================
  Hits         1382     1382           
  Misses         81       81           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[theogf]

Instead of the whole > v1.12, should we just drop Zygote support entirely? The tests are failing anyway.

[simsurace]

We could do that, but my plan was to have an intermediate release where extent of breakage/support is properly documented because I feel that not having a tested successor for AD is not a very good state to be in. So 0.11 would be a transition state. I feel that removing all Zygote hacks and making sure everything works with e.g. Mooncake or Enzyme is a much larger task, perhaps for a version v1? Enzyme for example has correctness bugs related to BLAS, I don't want to say "we dropped Zygote support on all versions, use something else" when that something is likely giving wrong results. Still something I would like to address in the near term, but there could potentially be a bigger step there, rewriting tests using ADInterface and testing different backends, perhaps also GPU support with AD is now finally within reach? At least it's worth investigating I think.

[theogf]

Yeah integrating DITests is something I want to get done at some point! Since it's just a temporary thing I guess it's fine to merge as is even if the _TEST_ZYGOTE pattern is not great.

[simsurace]

The plan would be to merge this and then release as 0.11. Would that be ok for you?

[github-actions]

KernelFunctions.jl documentation for PR #592 is available at:
https://JuliaGaussianProcesses.github.io/KernelFunctions.jl/previews/PR592/