Support for more general observational noise covariance in the joint likelihoods

[jlperla]

Right now we only support diagonal mean zero. For general MvNormal we could replace the current code with a linsolve. In fact, due to dispatching with diagonal matrices it might not require a special case.

See

DifferenceEquations.jl/src/linear.jl

Line 152 in 63a7cbe

Δz = Δlogpdf * (prob.observables[:, t_n] - z[t]) ./ abs2.(prob.obs_noise.σ) # More generally, it should be Σ^-1 * (z_obs - z)

for an example.