Arguments usage for general models such as AgeReplacementModel or regressions

[Nathan-Herzhaft]

The dynamic use of arguments like ar, a0 and covar make usage quite simple when chaining models. However, things get more complicated for new features developments because these attributes are hidden in params tuple arguments and are difficult to access.

General methods for models that might have an ar in the params or not are complicated beacause of this choice of implementation. Examples with the implem of sample for stochastic processes, or rvs with LeftTruncated + AgeReplacement.

We should think of a better way to make these optional arguments accessible, maybe dicts or other data structures to access them in an optional_arguments param.

Additionaly, covars present even more difficulty with a geometric shape that can vary depending on the number of assets, methods etc.

[Nathan-Herzhaft]

A nice user/dev - friendly implementation could be to join these arguments with the parameters arguments of parametric models. We would have a single arguments that would be a dict, containing entries for each parameter the model has to access.

For example, a Weibull could have .params = {"shape"=... , "rate"=...}, while AgeReplacement(Weibull) could have .params = {"shape"=... , "rate"=..., "ar"= ...}.

This implementation seems to be funcitonal for frozen and and unfrozen models. It would make easier the access of optional arguments in chained models, for users and devs.