ComposedEstimands shouldn't take a function

[olivierlabayle]

At the moment ComposedEstimands are defined by a function which means only one function can be applied to the joint. I think we should always take the joint distribution and later potentially evaluate any function on it.

!! Breaking (again)