Allow `prep=nothing` to stand for "no preparation"

[MilesCranmer]

Tiny tiny thing.... So, I have some internal methods that process and cache prep for a gradient call. Right now I have to do this at the call site:

(val, grad) = if isnothing(prep)
    value_and_gradient(f, backend, tree)
else
    value_and_gradient(f, prep, backend, tree)
end

I think I would enjoy if DifferentiationInterface.jl did this routing for me. If prep=nothing I think it should be semantically identical to not passing prep at all.

[gdalle]

Hi!
So in terms of code it's actually not so tiny, and I might want to do it at the same time as a large refactor I'm planning on how prep is handled. I agree that it would be useful though, and I have run into the same unsatisfying patterns.
Another option in the meantime is to have a tuple maybe_prep that is either (prep,) or (), and then do gradient(f, maybe_prep..., backend, x). Not very pretty but who knows, maybe you'll like it better.