Selectively reduce optimisation level in GHC build

Currently there are a few places in the GHC build where we do signficantly more time than is necessary or warranted, especially during development:

• the stage0 ghc-cabal is built with $SRC_HC_OPTS, which contains -O in most build flavours. This means that we end up spending a significant amount of time compiling an optimized Cabal, despite the fact that stage0 ghc-cabal is hardly a performance critical executable.

• stage2 Cabal is built with HsLibHcOpts, which contains -O in most build flavours. Not only is Cabal typically not going to be performance critical for users, but the object size from an unoptimised build is significantly smaller than that from an optimised build.