Cleaning up SVI-related internals

[eb8680]

This issue is for discussing instances of cruft or technical debt in SVI-related code and proposed fixes. We can create new issues for individual instances if needed, and add or remove items from this list as we decide what to focus on.

Potential issues/fixes:

• [ ] Improved support for nested SVI: see #1227
• [x] Collect all the duplicated *ELBO._get_traces logic into a single set of functions in infer.enum, as suggested in e.g. #1214. Ideally we could also use the same code as part of #1226 and #1128. (Issue moved from #1227 and work started in #1230)
• [ ] Remove deprecated infer.enum.iter_discrete_traces
• [ ] Remove SVI.evaluate_loss in favor of calling loss(...) directly
• [ ] Collect duplicated Trace*_ELBO.loss functionality into one place, e.g. ELBO.loss
• [ ] Split some infer.ELBO functionality into reusable functions that can be used outside SVI e.g. in Importance
• [ ] Standardize interfaces of SVI validation logic (e.g. check_model_guide_match, check_site_shape) so that it's more composable
• [x] Simplify ParamStore and PyroOptim by removing param_name() and related internals mapping from parameters to names
• [x] #1363 Make ParamStore more dict-like by replacing current interface with __getitem__/__setitem__/update, or even just make it a subclass of dict
• [x] #1307 Make iarange and irange look like more other side effects, i.e. make them into Messengers, including a functional interface
• [x] Possible: get rid of Trace data structure in favor of a simple OrderedDict

[fritzo]

To help guide this effort, I'd like to take a look at some long-term goals:

• [x] making newton_step() and Newton look more like SVI or TracePosterior
• [x] making MarginalAssignment and friends look more like SVI or TracePosterior
• [x] using SVI.differentiable_loss() inside newton_step() or Newton
• [x] #1128 supporting parallel enumeration in HMC
• [ ] #1213 batched params
• [ ] broadcasting semantics for EmpiricalMarginal