Summary:
This improves performance, at least sometimes--the previous
implementation can be worse than the version in base 4.7. I
have not had the time to run benchmarks and such, but `mapM`
already does this.

Also, inline `mapM_`, like `mapM`.

Reviewers: hvr, nomeata, ekmett, austin

Reviewed By: ekmett, austin

Subscribers: thomie

Projects: #ghc

Differential Revision: https://phabricator.haskell.org/D632

GHC Trac Issues: #10034