Relax rules for structured VMP in Transition node

[wouterwln]

With UnfactorizedData, we should be able to send PointMass messages with a factorized a as well. The rules are the same as for a Categorical (as far as I can see) but we didn't dispatch on them yet.

[ThijsvdLaar]

Does this change rely on an implementation of BayesBase.probvec(::PointMass)? Since the point mass rules are only supposed to work with one-hot encoded vectors, how is this handled?

[bvdmitri]

I think the current probvec simply returns the underlying vector without checking if its one hot encoded or not:

julia> using BayesBase

julia> probvec(PointMass(1))
ERROR: probvec(::PointMass{ <: Real }) is not defined
Stacktrace:
 [1] error(s::String)
   @ Base ./error.jl:35
 [2] probvec(::PointMass{Int64})
   @ BayesBase ~/.julia/dev/BayesBase/src/densities/pointmass.jl:68
 [3] top-level scope
   @ REPL[2]:1

julia> probvec(PointMass([1, 2]))
2-element Vector{Int64}:
 1
 2

[ThijsvdLaar]

Ok, seems reasonable to not check for one-hot encoding, maybe make a note of this for clarity.