Introduce type hierarchy for extras

[gdalle]

This PR is motivated by a clean preparation of the HVP operator. It ended up leading to a revamp of the way extras work. The HVP preparation is still wasteful but improving it will be much easier now.

Core

• BREAKING: Remove support for scalar input in the gradient and HVP operators

• Define abstract supertypes for each operator: PushforwardExtras, GradientExtras, etc.

• Define corresponding empty structs: NoPushforwardExtras(), NoGradientExtras(), etc.

• Ensure that each operator only accepts extras of the right type

• Define sensible default extras for higher-level operators

struct PushforwardJacobianExtras{E<:PushforwardExtras} <: JacobianExtras
    pushforward_extras::E
end

struct PullbackJacobianExtras{E<:PullbackExtras} <: JacobianExtras
    pullback_extras::E
end

• Define conversion mechanisms to avoid recomputing extras in lower-level operators:

function prepare_pushforward(
    extras::PushforwardJacobianExtras, f_or_f!, backend::AbstractADType, args...
)
    return extras.pushforward_extras
end

function prepare_pullback(
    extras::PullbackJacobianExtras, f_or_f!, backend::AbstractADType, args...
)
    return extras.pullback_extras
end

Extensions

• Use No...Extras in lieu of extras::Nothing
• Define custom structs storing configs and tapes whenever necessary

struct ForwardDiffGradientExtras{C} <: GradientExtras
    config::C
end

struct ReverseDiffGradientExtras{T} <: GradientExtras
    tape::T
end

Test

• Remove tests for scalar input in the gradient and HVP operators
• Adjust zero backends