Fix, benchmark and re-evaluate -fexpose-all-unfoldings and -fspecialise-aggressively

These had to be removed, because they caused OOM on armv7.

They bit-rotted (the Cabal optimization flags were probably not moved to the solver package when it was separated out), so the benchmarks in https://github.com/haskell/cabal/pull/7729#issuecomment-939052965 show they have negligible effect.

The `-fexpose-all-unfoldings` for `parsec` is a somewhat different story, but it needs re-evaluation as well, ideally sooner rather than later (low cost of slower compilation vs potentially higher gain). These benchmarks are probably still valid: https://github.com/haskell/cabal/commit/dcb23dd33fb3a76b4ab98b3f2100f851bb1d447e

I think, before 3.8 or after GHC 9.2 is out, whichever comes first, it makes sense to fix the optimizations and re-benchmark and then decide.