Try parametric dispatch for performance optimisations

[SamuelBrand1]

          > Based on the benchmarks this looks like its quite a lot slower for all low-level models (and dependent benchmarks) in all cases except where compiled reverse diff is used. I think that is probably just the cost of dispatching through the multiple new submodel calls vs specifying the iid process directly and so somewhat unavoidable if we want this functionality (which I think we do)

This seems possible. I've spotted some places we could have tried a parameteric dispatch e.g. have a method where the $\epsilon_t$ field of IID being a concrete Normal type dispatches to using MvNormal

Originally posted by @SamuelBrand1 in #438 (comment)

As described above. The basic code pattern would be something like

@model function generate_latent(IID{D},...) where D <: Distribution
	...opt code for D...
end