We were previously doing this during extraction, i.e. for each field
node we encouter, which meant we were repeating a lot of work. The
`type_matches_set` function was a fairly significant hot-spot in
profiling results, so this should improve performance.

First, emit labels with fresh ids. Then other labels. Then tuples,
grouped by name. Hopefully this will help both with the compression
ratio but also with branch prediction in the TRAP importer.

This is the same idea as Java's LinkedHashMap: it gives the same O(1)
insertion and lookup as HashMap, but preserves insertion order for
iteration.