Consider replacing NRT annotation skill with a slimmer version

See experiment writeup in https://gist.github.com/danmoseley/2bd583fb655c9248828c0f8b339ac6c2

We should probably take something like the "wisdom" variant that is a heavily cut down variety of the checked in one and run it on https://github.com/DapperLib/Dapper. Basically check out the ref just before Marc's big change to NRT annotate dapper (which is considered "gold standard" for this experiment) and run both current skill, trimmed skill, and possibly no skill against that commit to NRT annotate it. Compare to gold. If the trimmed skill performs better, replace it.