Implement AutoGaussian guide

[fritzo]

Addresses #2813
Replaces #2926 #2928
See design doc

Tasks

• Parametrize sparse precision matrices. This is statistically optimal.
• Implement "dense" backend. This is fully general but computationally expensive.
• Implement "funsor" backend. This is asymptotically cheaper than the dense backend but has two restrictions (from TVE):
  • For any dependency u --> d, plates(u) <= plates(d).
  • For any pair of transitive dependencies u1 --> d <-- u2, either plates(u1) <= plates(u2) <= plates(d) or plates(u2) <= plates(u1) <= plates(d).

After this first implementation, I plan to switch to forward_filter_backward_precondition() and either compile or trace the funsor computation to reduce overhead.

Tested

• added to existing autoguide tests
• smoke test on pyro-cov model variants
• test equality of elbo to AutoMultivariateNormal (on a gaussian model)
• test on lots of dependency structures

[eb8680]

Looks good per zoom review pending updated serialization test

[fritzo]

OK, I've added serialization tests and fixed a few serialization issues they exposed.