Merge the hs-venture branch

[axch]

This is the branch where the Haskell Venture backend lives. While it is very far from feature-complete with respect to Lite or Puma, I think it's appropriate to stop hiding it on a branch. Among other reasons, my memory is that it runs the curve fitting demo visually faster than Puma does (presumably because the web server is in Haskell rather than Python and there is far less stack goo between it and the actual computation). I am also tired of switching branches to see how I did something in it, or show someone else how I did something in it.

The main impediment is the web demos: I edited them slightly on the hs-venture branch, primarily to make them conform to a stricter wire protocol (server-side DWIMs being more seductive in Python than Haskell).

• [ ] Those changes would need to be thoughtfully merged.

Additional things that might be reasonable to do if touching that code anyway:

• [ ] Move to using stack as the build system for HsVenture, rather than raw cabal
• [ ] Add a Jenkins job that runs HsVenture's test suite
• [ ] Debug inference quality in the crp clustering demo (presumably only the uncollapsed version, because I don't want to block on implementing any collapsed normal models in HsVenture)

[axch]

I started HsVenture in order to have written my own Venture implementation. Aside from intellectual interest, it does have one capability we might actually want to use: Haskell has a trustworthy library for automatic differentiation, so HsVenture has the fastest and least buggy gradients of all extant Venture implementations (and can be a playpen for methods that need higher order derivatives, like Riemannian Manifold Hamiltonian Monte Carlo).

[axch]

git diff --stat origin/master...origin/hs-venture tells me that changes are confined to the backend/hs directory and to couple web demos in demos/jsripl.

[axch]

Fixed by commit fd6c12e45b2d213312117fa40360f77a2390889e.