Add the option couple to AdamW and set the default to match pytorch

[CarloLucibello]

See FluxML/Flux.jl#2433 for the details

Close FluxML/Flux.jl#2433

[ToucheSir]

Suggested change

	function AdamW(η, β = (0.9, 0.999), λ = 0.0, ϵ = 1e-8; couple::Bool = true)
	function AdamW(η, β = (0.9, 0.999), λ = 0.0, ϵ = 1e-8; couple::Union{Nothing,Bool} = nothing)

Could we do something like this and then add a warning in the constructor if couple isn't set?

[mcabbott]

I'm happy to change the default but I wonder if this couple keyword is too easy to miss, and too obscure a name. If you read someone's code, would you ever you guess what it means?

One extreme would be to have AdamWpaper and AdamWtorch, then neither gets the neutral name.

[CarloLucibello]

I would leave this PR as it is. 99% of users will write AdamW(lr) and get what pytorch does and will be happy. We want couple to just not appear in most codebases and not create mental burden to most users. It will be available and clearly discoverable through the docstring for the few people aware of the paper/pytorch discussion.