Greatly speed up the import-cache computation.
If the imported context does not build a loop with the base one, just efficiently merge its imported set, instead of recursing through all imports.

This took about 20% of parsing time, and now should hopefully be near zero.