Do not claim that -O2 does not do better than -O

when in fact it does. This was pointed out by Johannes Bechberger and
supported with seemingly statistically sound evidence in his Bachelor
thesis: Of the benchmark shootout programs, 80% benefit significantly by
switchtng from -O to -O2.

See https://uqudy.serpens.uberspace.de/blog/2016/02/08/ghc-performance-over-time/
for a few raw numbers.

Differential Revision: https://phabricator.haskell.org/D2065

(cherry picked from commit 3d245bf5255ebfb72813596fa93b9051f7518321)