I was alerted to this by Trac #12468 and #11325.  We were treating
insolubles (and "hole" constraints are treated as insoluble)
inconsistently.  In some places we were carefully rewriting them
e.g. Note [Make sure that insolubles are fully rewritten] in
TcCanonical.  But in TcSimplify we weren't feeding them into
the solver.

As a result, "hole" constraints were not being rewritten, which
some users found confusing, and I think rightly so.

This patch also fixes a bug in TcSMonad.emitInsoluble, in which two
different "hole" constriants could be treated (bogusly) as duplicates,
thereby losing one.