Test SVI for a model where we have NUTS results

[SamuelBrand1]

From f2f conversation there was some interest in exploring variational inference; mainly to resolve long compute times.

My understanding is that the infrastructure exists in the code to swap samplers and a variational inference sampler exists for numpyro https://num.pyro.ai/en/stable/svi.html .

So in principle, for situations where we already have NUTS generated posterior samples, it should be quite easy to compare the results.