Follow-up from "Fix the binder-swap transformation in OccurAnal"

The following discussion from !5351 (closed) should be addressed:

• @sgraf812 started a discussion: (+1 comment)

  This is all very similar to GHC.Types.Unique.SDFM.UniqSDFM, except that

  1. UniqSDFM is deterministic, which means we lose performance if we switch to it
  2. UniqSDFM isn't concerned about coercions at the moment

  (1) is easily rectified by extracting an internal type class for emptyUFM, lookupUFM and extendUFM so that we can share the implementations.

  (2) is something that the pattern-match checker should have done for a long time anyway; currently we just discard coercions, which is a bug waiting to happen.