Use full mass matrix for globals in SIR example

[fritzo]

Addresses #2426

This updates the SIR example to:

1. Use a full mass matrix for global parameters R0,tau,rho
2. Add a command line arg to ensure a sufficient number of observations are generated. This is important because some parameters are not identifiable in the limit of small observations / population.

Anecdotal Evidence

Before this PR:

                mean       std    median      5.0%     95.0%     n_eff     r_hat
        R0      1.59      0.07      1.59      1.49      1.72     10.96      1.20
       tau      6.84      0.44      6.85      6.18      7.52      5.06      1.83
       rho      0.48      0.02      0.48      0.45      0.51      7.80      1.22

After this PR:

                mean       std    median      5.0%     95.0%     n_eff     r_hat
        R0      1.54      0.07      1.53      1.44      1.67     12.11      1.32
       tau      6.41      0.46      6.35      5.69      7.18      6.88      1.51
       rho      0.49      0.02      0.49      0.46      0.52     10.34      1.00

Tested

• existing smoke tests
• ran locally with --population=10000 --min-observations=1000

[martinjankowiak]

fyi: travis is failing with ValueError: Failed to generate 3 observations. Try increasing --population or decreasing --min-observations FAILED