Verify that we get 100% precision and recall for equations

[Chris Warburton]

Let's say we extract all theorems from the TIP benchmarks and discard
those which aren't equations; call the resulting equations 'eqs'. We
should be able to convert eqs into JSON, and feed it into the
precision/recall checker with eqs as the ground truth, resulting in 100%
precision and recall.

I've added a test for this in conjectures.rkt, but it currently only
gives 97% precision....