Issue 176: Add support for multiple observation models

[seabbs]

This PR closes #176 by adding a new StackObservationModels which supports arbitary stacks of observation models with both 1 to 1 and 1 to many infection -> obs mappings.

Note that this does not support the passing of unnamed y_t in the generated quantities (in the sense of it making no effort to track these) as long term this doesn't seem sustainable - especially due to the lack of integration with prefix.

This implies we are expecting posterior predictions to be done on data using the y_t (with prefix) sampled at the lowest level of the stack.

if we did want to keep more in the returns this we would need manually prefix I think. To me this suggests a package wide rethink of the use of return is needed.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 93.34%. Comparing base (599c32b) to head (231d9eb).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #249      +/-   ##
==========================================
+ Coverage   92.97%   93.34%   +0.37%     
==========================================
  Files          43       44       +1     
  Lines         413      436      +23     
==========================================
+ Hits          384      407      +23     
  Misses         29       29              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[seabbs]

Implementing this has also made me thing about hierarchical priors. Currently the only way I can think we might do this is similar to this stacking approach but passing in uninstantiated structs and a prior model function and then in the method instantiating the structs using the prior model.

This feels a bit clunky? As far as I am aware we don't have an issue for this so shall we split out and discuss? A good test case for this feels like it might be trying to fit to counts and deaths with a partially pooled overdispersion parameter. Another sensible toy could be trying to fit to cases and deaths with both having 3 age groups (and pooling by dataset and age group both nested and independently).

Another option might be allowing structs to be partially instantiated but that feels very bug prone

[SamuelBrand1]

I like this! This is really neat.