Test avoiding InExactError by using TransformObservationModel

[SamuelBrand1]

As a reminder this is caused by the inference expecting Int valued returns but instead getting caught by a BigInt. This is possible with latent processes in the log domain which gets exp transformed and then linked with a neg binomial or Poisson distribution e.g.

julia> typemax(Int) |> log
# 43.66827237527655

We've tried a few approaches to stop this eg #492 .

This issue is about trying a TransformObservationModel composition with the usual obs processes to implement:

$$x \to - \log ( e^{-C} + e^{-x})$$

Where $C$ is a large float. In julia this is

using LogExpFunctions
const C = some_big_number
x -> -logaddexp(-C, -x)

[SamuelBrand1]

The point is that its smooth at the transition point but also effectively linear for most of the domain of relevance.

[SamuelBrand1]

Closed by #562