Interface tweaks

[null-a]

We might be able to give some things better names:

• iter parameter of nuts to num_samples
• fit.fitted to fit.predict? (both exist in brms, and IIRC fitted in brmp implements aspects of each of these.) Alternatively, we could move away from a single mega-method towards separate methods for each kind of prediction? (i.e. We could have separate methods, one for each valid what argument that can be passed to fitted.)
• defm to brm? (defm is short for "define model", but it wasn't a well considered choice. brm is used in brms for something similar but not quite the same -- it both defines and fits a model -- which might cause confusion.)
• DefmResult and GenerateResult might seem a bit odd when they appear in documentation.
• Drop the nuts,svi,prior methods on DefmResults? (i.e. Remove support for doing defm(...).prior().) These are a middle ground between the convenience of def(...).fit()and the flexibility of e.g. defm(...).generate(...).nuts(...), but it's not clear they add much. Also, they are currently hard-coded to use only the default back end.