Support for combining models

[richfitz]

This PR adds support for combining models via adding their log densities. The basic idea is to support the fairly common case that we have a pair of models, one representing the (log) priors and the other representing the (log) likelihood that we want to join together to make a posterior. We can then construct these however we want and combine them with mcstate_model_combine (or + which is syntactic sugar for this).

The way that different model components combine is described in some detail in the docs but briefly:

• parameters become the union over both models
• domains take the most restrictive per parameter
• densities are summed on the log scale
• gradients are also summed on the log scale, assuming (correctly) zero gradient for variables missing from a component model
• direct sample can only be included from a single model

Because I was finding it hard to work with, I've changed domain slightly to include parameter names.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 100.00%. Comparing base (06322fa) to head (08c60e3).

❗ Current head 08c60e3 differs from pull request most recent head 5fbb465. Consider uploading reports for the commit 5fbb465 to get more accurate results

Additional details and impacted files

@@            Coverage Diff            @@
##              main       #14   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files           39        40    +1     
  Lines         1689      1689           
=========================================
  Hits          1689      1689           

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

[M-Kusumgar]

Nice one, was pretty clear, couple small suggestions!

[M-Kusumgar]

Suggested change

	## Next we need to check that the model with a sample has the full set of
	## Next we need to check that the model with a sample has the full set of parameters

[M-Kusumgar]

im not sure that the name "a" or "b" is useful to the user, i found it actively confusing in the test for this, can we do a

`+.mcstate_model` <- function(x, y,
  names = c(deparse(substitute(x)),deparse(substitute(y))))

and pass this down into this function so that we can reference the variable that they tried to sum?

[weshinsley]

Looks neat - some tiny doc typos!

[M-Kusumgar]

Nice one, i like the "lhs" and "rhs", much clearer!