On a test failure, we show a diff between the expected and the actual
output. The method of how we do this has changed a couple of times:

* In 2007: 9951189c
  "On failure, diff the normalised test outputs"

* In 2011: 3019b1e4
  "When the output files differ, present the diffs between the *actual*
  output, not the normalised output. The latter may have newlines
  removed, making the diff unreadable."

* In 2015 (now): do something in between.
  - Do apply the normalisers again, to make the diff smaller (only
    showing the actual problem).
  - But don't apply normalise_whitespace, as it indeed makes the diff
    unreadable.

Differential Revision: https://phabricator.haskell.org/D984