Add gradient preparation to invalidation rate

[pat-alt]

@gdalle I've tried what you suggested in #518, but the prepared gradient doesn't match the executed gradient:

/home/runner/work/CounterfactualExplanations.jl/CounterfactualExplanations.jl/test/generators/probe.jl:19
  Got exception outside of a @test
  PreparationMismatchError (inconsistent types between preparation and execution):
    - f: ✅
    - backend: ✅
    - x: ❌
      - prep: Matrix{Float32}
      - exec: Matrix{ForwardDiff.Dual{ForwardDiff.Tag{typeof(CounterfactualExplanations.Generators.pen_wrt_state), Float32}, Float32, 2}}
    - contexts: ✅
  If you are confident that this check is superfluous, you can disable it by running preparation with the keyword argument `strict=Val(false)` inside DifferentiationInterface.

Any idea what be going on? Note that this gradient computation is nested inside another gradient computation - perhaps that's the issue?

[gdalle]

Any idea what be going on? Note that this gradient computation is nested inside another gradient computation - perhaps that's the issue?

Your diagnosis is correct, I think you might need to replace similar(ce_state) in your preparation step by an array of the correct dual eltype. It's one of the most cumbersome things about DI, and functions like overloaded_input_type may help you do that, but an MWE would be most useful.