Tape-able pullback with ReverseDiff

[gdalle]

At the moment, the ReverseDiff pullback is a bit hacky because it takes the gradient of a closure. Ideally we should hook into ReverseDiff internals

[gdalle]

Essentially, the idea is to use

• prepare_pullback to prepare the tape
• pullback to execute it and get the tangent

The issue is that the tape is a ReverseDiff internal, and functions with array outputs are not even supposed to be supported (but apparently they are if we go deep enough?).

There is nothing similar in other DI extensions, the closest thing might be symbolic backends, where I use the preparation to generate a fast executable: see e.g.

• FastDifferentiation with 1-argument functions: https://github.com/gdalle/DifferentiationInterface.jl/blob/73a2ddd9e86f9525f35abf369c2e58ab4783509b/DifferentiationInterface/ext/DifferentiationInterfaceFastDifferentiationExt/onearg.jl#L71-L132
• FastDifferentiation with 2-argument functions: https://github.com/gdalle/DifferentiationInterface.jl/blob/73a2ddd9e86f9525f35abf369c2e58ab4783509b/DifferentiationInterface/ext/DifferentiationInterfaceFastDifferentiationExt/twoarg.jl#L81-L155

These are exactly the operators we need to (re)implement in the ReverseDiff extension

[gdalle]

Not worth wasting sweat on ReverseDiff now that Mooncake is here