Refactoring integration tests

[neerajprad]

Remaining action item from #120:

Refactor the existing integration tests to inherit from a common structure - where the model, guide, inference algorithms and their parameters, expected posterior etc. can be specified via parameters. New inference algorithms can reuse this structure.

• [ ] To begin with, have some common tests for MCMC algorithms (no guide specification needed). Pull in some tests from webPPL.
• [ ] Unify the current inference, tracegraph, map_data, and sampling integration tests, and generate them automatically - there is a danger of over-abstraction and resulting inflexibility which we need to be careful about, since the whole idea is to make it easy to test new inference algorithms or improvements to existing ones.

[neerajprad]

I think we have a lot of tests for HMC and NUTS already. We can reopen this if we decide to implement MH. Regarding (2), I think we have refactored many of our tests and made them much faster, so let us have a more targeted issue if we want to invest more time in refactoring and unifying these tests.