Move to all Turing models in EpiAware to be non-vectorised

[SamuelBrand1]

The problem with a generic forecast function for EpiAware #243 is that in general the Turing.forecast function doesn't interface smoothly with vectorised random variable sampling (see this discussion TuringLang/Turing.jl#2239).

In various f2f discussions (including with @dylanhmorris ) its been noted that these issues could (probably; we'd need to do some dev work) be resolved by going to a non-vectorised approach (e.g. filldist with Normal r.v.s rather than MvNormal). The downside is the potential performance hit (e.g. https://turinglang.org/docs/tutorials/docs-13-using-turing-performance-tips/ ).

If we want to do this on the way to resolving #243. I think a good step would be to first resolve #300 so we have an accurate before and after picture of performance.

[seabbs]

its been noted that these issues could (probably; we'd need to do some dev work) be resolved by going to a non-vectorised approach

I agree this seems like it should be better supported with current functionality.

I also agree this is blocked by benchmarking which I will work on asap so we can unstick this

[SamuelBrand1]

Is this unblocked now?

[seabbs]

I think we want the full benchmarking and integration profiling I am doing tomorrow in place first