Use SMC to initialize compartmental models

[fritzo]

Addresses #2426

This replaces custom .heuristic() methods with a default generic SMCFilter-based initialization strategy, greatly simplifying the process of creating new compartmental models. Users can still override this default .heuristic() if needed.

🎉 hooray for composable inference 🎉

Tested

• refactoring covered by existing tests
• ran python sir.py -p 10000 -d 60 -f 30 --verbose --plot and confirmed good posteriors and mixing:

         mean       std    median      5.0%     95.0%     n_eff     r_hat
 R0      1.48      0.08      1.47      1.36      1.61     16.89      1.09
rho      0.48      0.03      0.49      0.44      0.53     29.28      1.14

[martinjankowiak]

why is this needed exactly?

[fritzo]

Recall that SMCFilter tracks all state in a state dict and periodically resamples all that state. In this use case the state dict contains values of all sample sites that we will eventually replay to initialize MCMC. In particular it contains sample sites for global variables. Now since resampling changes those sites, and since global_model() not only draws samples but performs computation on those samples, we need to re-perform that computation after each resampling. Hence we block (to avoid this being traced again), and condition on the SMCFilter-managed state to get the latest resampled params each step.