Provide standard API to get **one sample** in a "structured" way

Hey,
This might be me wanting crazy things again but here is my proposal:
Should we provide an abstract function (to overload), taking state, sample and model to return samples in a predefined way.
Let me give an example. Given a Turing model

@model function mymodel(x, y)
	s ~ Gamma(1, 1)
	w ~ Normal(0, s)
    y ~ MvNormal(x .* w, 1)
end
x = rand(5)
y = rand(5)

there would be a function which given the current sample and model returns (s=0.354123, w=0.154658)

The motivation behind this is to make more "universal" callback functions (the callback mafia is here again). Right now defining a callback function needs to be tailored for each model.
However having a to_named_tuple (which would be optional to implement), would allow to build many more generic functions built upon AbstractMCMC.

So I am pretty sure I am missing out why this is not making sense or that I am missing the bigger picture and I would be happy to hear why I am terribly wrong about this :)

PS: The NamedTuple is just a proposition, my proposal is more about having a convention about the output.

Review the existing AbstractMCMC callback conventions and the proposed inputs of state, sample, and model. The issue is complete only when the project agrees on a structured single-sample API, including its output convention and implementation scope.