This commit fixes some performance regressions introduced by 0cc47eb9,
adding more `Coercible` magic to the solver. See Note
[flatten_many performance] in TcFlatten for more info.

The improvements do not quite restore the old numbers. Given that
the solver is really more involved now, I am accepting this regression.

The way forward (I believe) would be to have *two* flatteners: one
that deals only with nominal equalities and thus never checks roles,
and the more general one. A nice design of keeping this performant
without duplicating code eludes me, but someone else is welcome
to take a stab.