Learn white_vec parameter of AutoGaussian guide

[fritzo]

Addresses #2813
Blocked by pyro-ppl/funsor#553

This adds learnable parameters white_vec to each Gaussian factor in the AutoGaussian guide, so the total set of factors is: loc, scale, white_vec, and prec_sqrt. The parameters (white_vec, prec_sqrt) are important because they are closer to the natural parameters than (loc, scale), and should in theory lead to faster learning. The scale parameter is important because (i) being on a log scale it can quickly adapt to large or small magnitudes, and (ii) it performs row-size scaling of the precision matrix and leads to faster learning (see #2127, #2924). It appears that the loc parameter is also required.

Tested

• strengthened a dispatch test
• pass existing tests

[fritzo]

@eb8680 this isn't super high priority, but it would be nice to get it in for the first release of AutoGaussian. Nothing has changed since we discussed this a month or two ago.

[eb8680]

Looks good, merging