Doing canonicalization on every comparison turned
out to be very expensive.

Caching the canonicalization through the smart `modulePair` constructor
gives `8%` reduction in allocations on `haddock.compiler` and
`8.5%` reduction in allocations on `haddock.Cabal`.
Possibly other things as well, but it's really visible in Haddock.

Test Plan: ./validate

Reviewers: jstolarek, simonpj, austin, simonmar, bgamari

Reviewed By: simonpj, simonmar

Subscribers: thomie

Differential Revision: https://phabricator.haskell.org/D2350

GHC Trac Issues: #12191