Feat: A third preparation mode between `strict=Val(true)` and `strict=Val(false)`: fall back to unprepared execution on signature mismatch

Suggested by @gdalle in rsenne/ParallelMCMC.jl#62

the strict kwarg in prepare_gradient has two modes: true and false with the basic idea that true just ensures the type matches else throws PreparationMismatchError

It could be useful to have a middle ground between the lax and strict mode (In ParallelMCMC I have a
gradient prepared for Vector{Float64} that gets called on Vector{<:Dual}). E.g., like i wrote in PMCMC:

struct _ADGradient{F,B<:AbstractADType,P,TX}
    logdensity::F
    backend::B
    prep::P
end

function (g::_ADGradient{F,B,P,TX})(x) where {F,B,P,TX}
    if x isa TX
        return DI.gradient(g.logdensity, g.prep, g.backend, x)   # fast path
    else
        return DI.gradient(g.logdensity, g.backend, x)           # from scratch
    end
end

So simply, if prep matches, great, else run from scratch

Start by reading prepare_gradient and its existing strict=true and strict=false behavior, then use the proposed _ADGradient callable path as the concrete reference for the fallback semantics. Done means a third preparation mode is defined and implemented so matching signatures use the preparation while mismatches execute from scratch.