particles failing on some For distributions

[cscherrer]

@DilumAluthge saw a problem in cscherrer/SossMLJ.jl#110 that seems to have its root in Soss.particles:

julia> Soss.particles(For(identity, Normal.(randn(3), randn(3).^2)))
ERROR: MethodError: no method matching Particles(::Int64, ::For{typeof(identity),Array{Normal{Float64},1},Normal{Float64},Float64})
Closest candidates are:
  Particles(::Integer) at /home/chad/.julia/packages/MonteCarloMeasurements/ETaFe/src/types.jl:71
  Particles(::Integer, ::Distribution{Multivariate,S} where S<:ValueSupport) at /home/chad/.julia/packages/MonteCarloMeasurements/ETaFe/src/types.jl:84
  Particles(::Integer, ::Distribution{var"#s47",VS} where var"#s47"; kwargs...) where VS at /home/chad/.julia/packages/MonteCarloMeasurements/ETaFe/src/types.jl:71
  ...
Stacktrace:
 [1] particles(::For{typeof(identity),Array{Normal{Float64},1},Normal{Float64},Float64}, ::Int64) at /home/chad/.julia/packages/Soss/I6MSl/src/particles.jl:12 (repeats 2 times)
 [2] top-level scope at REPL[47]:1
 [3] include_string(::Function, ::Module, ::String, ::String) at ./loading.jl:1088

[cscherrer]

Also, for background on the weird Particles hack we're currently using in Soss, baggepinnen/MonteCarloMeasurements.jl#22

[cscherrer]

When a model contains a For, the generated code calls the helper function parts, which in this case is

parts(d::For, N::Int=DEFAULT_SAMPLE_SIZE) = parts.(d.f.(d.θ...), N)

Changing this to parts.(d.f.(d.θ), N) fixes the problem for this case, but may break other uses of For.

A good way around this could be to add a collect(::For) method that returns a Vector{Distribution}. Then it's (hopefully) easy to make this do the right thing for any instance