Trapping constraints

[Jacob-Stevens-Haas]

I wanted to share this work as a draft that adds _make_constraints() from example 8 to trapping_sr3.py. All the differences are noted in the commit messages, but the big ones are:

• Term ordering is controlled by PolynomialLibrary.powers_.
• Creates constraints in a 3D numpy array, rather than 2D. Caller can flatten to either "target" or "feature" constraint order. IMO this makes the logic a lot clearer, since the term in $Q_{ijk}$ is explicit. i.e. there's no more long stride calculations like n_tgts * (n_tgts + k - 1) + i + n_tgts * int(j * (2 * n_tgts - j - 3) / 2.0).

Also a draft because I enabled too many PolynomialLibrary options. include_interaction=False would be incompatible with the trapping constraints IIUC. OTOH, connecting to powers_ gives us include_bias=True and interaction_only=True constraints for free.

[akaptano]

@Jacob-Stevens-Haas or @MPeng5 Have you checked the updated trapping constraints are consistent with @MPeng5 's branch? Mai, it might be worth merging your branch into this pull request and then rewriting your examples to be consistent with the new syntax here.

[Jacob-Stevens-Haas]

Oh damn, sorry, I thought I responded a while ago. TL; DR they should be consistent, but mine aren't int target/feature format - user has to flatten them to whichever format they want (there is code in the docstring for each type).

I'll check when I get back from break.

[Jacob-Stevens-Haas]

Note to self:

• pre-empted by work on #459
• Tests fail because using post-3.8 typing syntax.