Summary:
The algorithm for ApplicativeDo rearrangement is based on a heuristic
that runs in O(n^2).  This patch adds the optimal algorithm, which is
O(n^3), selected by a flag (-foptimal-applicative-do).  It finds better
solutions in a small number of cases (about 2% of the cases where
ApplicativeDo makes a difference), but it can be very slow for large do
expressions.  I'm mainly adding it for experimental reasons.

ToDo: user guide docs

Test Plan: validate

Reviewers: simonpj, bgamari, austin, niteria, erikd

Subscribers: thomie

Differential Revision: https://phabricator.haskell.org/D1969