Cleanup interface for the D

[jlperla]

After #100 we can cleanup the interface for the diagonal matrix.

• [ ] While we are at this, we probably also want to get rid of the TuringDiagMvNormal distribution, which also serves no purpose and was a hack prior to custom rules. It is possible that the only change required in DSSM is https://github.com/HighDimensionalEconLab/DifferentiableStateSpaceModels.jl/blob/main/src/types.jl#L368 into

  maybe_diagonal(x::AbstractVector) = MvNormal(Diagonal(x))

• [ ] In DifferenceEquations we then just need to access the diagonal a little different. In particular, instead of https://github.com/SciML/DifferenceEquations.jl/blob/main/src/linear.jl#L152 we want to store this above the for loop, e.g.

  cov_D = Diagonal(cov(prob.obs_noise)) # I think?  I don't think we want to square it in this form, but worth checking

  Then in that line, it becomes

  Δz = Δlogpdf * (prob.observables[:, t_n] - z[t]) ./ cov_D

  (or something along those lines. The key is that it looks like the covariance matrix of the ZeroMeanDiagNormal is not stored as a diagonal.

[jlperla]

And just to verify, it is important to create this as a MvNormal with a diagonal.

julia> MvNormal(Diagonal([0.1, 0.2])) |> typeof
ZeroMeanDiagNormal{Tuple{Base.OneTo{Int64}}} (alias for MvNormal{Float64, PDMats.PDiagMat{Float64, Array{Float64, 1}}, FillArrays.Zeros{Float64, 1, Tuple{Base.OneTo{Int64}}}})

Which has a very specialized type. If we wanted to backprop, then those tangent types should be easy as well.