Compute the union of imp_finsts on the side

I've explained most of the rationale in a new Note.
I'd happily add a test for this, but the difference is only
visible in run time, allocations remain more or less the same.

FWIW running `generateModules` from #14693 with DEPTH=16, WIDTH=30
finishes in `23s` before, and `11s` after.

Test Plan: ./validate

Reviewers: simonpj, simonmar, bgamari

Reviewed By: simonpj

Subscribers: rwbarton, thomie, carter

GHC Trac Issues: #14693

Differential Revision: https://phabricator.haskell.org/D4369