Posterior module API

[hmd101]

Separation of concerns: ParameterPosterior vs PredictivePosterior:

• ParameterPosterior: Its only job is to safely encapsulate the result of an inference engine (a point estimate for MAP, or eventually an array of samples for MCMC). Think of it as a data container.

• PredictivePosterior: takes data contained in ParameterPosterior and does the 'math' to e.g.,  predict_thresholds once, rather than re-implementing it inside MAPPosterior, LangevinPosterior, etc.

Replacing class inheritance with interfaces (Protocols):

• If an object has .params, .model, and .sample(), it is a ParameterPosterior.
• example: the user calls optimizer.fit(model, data). They don't need to know or care that they are getting a MAPPosterior. They just know the factory handed them back an object that fully satisfies the ParameterPosterior protocol. This makes the frontend API identical whether you're using MAP or MCMC.

Related to #87 and shelving #87 for now as the current API separating Model (prior and likelihood) from Posterior is in line with other established Bayesian libraries.