This regrettably-big patch substantially improves the way in which
"improvement" happens in the constraint solver.  It was triggered by
trying to crack Trac #10009, but it turned out to solve #10340 as
well.

The big picture, with several of the trickiest examples, is described
in Note [The improvement story] in TcInteract.

The major change is this:

* After solving we explicitly try "improvement", by
     - making the unsolved Wanteds into Deriveds
     - allowing Deriveds to rewrite Deriveds
  This more aggressive rewriting "unlocks" some extra
  guess-free unifications.

* The main loop is in TcInteract.solveSimpleWanteds, but I also ended
  up refactoring TcSimplify.simpl_loop, and its surrounding code.

  Notably, any insolubles from the Givens are pulled out
  and treated separately, rather than staying in the inert set
  during the solveSimpleWanteds loop.

There are a lot of follow-on changes

* Do not emit generate Derived improvements from Wanteds.
  This saves work in the common case where they aren't needed.

* For improvement we should really do type-class reduction on Derived
  constraints in doTopReactDict.  That entailed changing the GenInst
  constructor a bit; a local and minor change

* Some annoying faffing about with dropping derived constraints;
  see dropDerivedWC, dropDerivedSimples, dropDerivedInsols,
  and their Notes.

* Some substantial refactoring in TcErrors.reportWanteds.
  This work wasn't strictly forced, but I got sucked into it.
  All the changes are in TcErrors.

* Use TcS.unifyTyVar consistently, rather than setWantedTyBind,
  so that unifications are properly tracked.

* Refactoring around solveWantedsTcM, solveWantedsAndDrop.
  They previously guaranteed a zonked result, but it's more
  straightforward for clients to zonk.