Should Flux.destructure be moved to Functors.jl

[theogf]

I absolutely love the functionality of destructure, however I am surprised it is existing in Flux.jl instead of Functors.jl since it is not "Neural Network specific".
Maybe it is worth moving the function to Functors?

[ToucheSir]

Looks like destructure relies on Zygote.Buffer, so it might be difficult to port over without also taking on that dependency. On the bright side, the definition is pretty short and requires no additional dependencies if one uses a normal vector instead of a Buffer.

[mcabbott]

#31 and FluxML/Optimisers.jl#40 propose that:

• Functors should have some functions for making a flat array of all numeric contents, and
• Optimisers should have the direct replacement for destructure, which includes only trainable parameters.

The tricky part is getting the gradient correct & robust. An open question is whether Functors wants to depend on ChainRules to define this, or leave that part to Optimisers (which already depends on CR).

[ToucheSir]

destructure is now officially part of Optimisers.jl.