Before, it was possible to have a datatypes such as

```
type ConstantT a b = a
newtype T (f :: * -> *) (a :: ConstantT * f) = T (f a) deriving Functor

data family TFam (f :: * -> *) (a :: *)
newtype instance TFam f (ConstantT a f) = TFam (f a) deriving Functor
```

fail to eta-reduce because either (1) a TyVar had a kind synonym that
mentioned another TyVar, or (2) an instantiated type was itself a type
synonym that mentioned another TyVar. A little bit of tweaking to
`expandTypeSynonyms` and applying it before the eta-reduction check in
the `deriving` machinery is sufficient to fix this.

Fixes #11416.

Test Plan: ./validate

Reviewers: goldfire, simonpj, austin, bgamari

Reviewed By: simonpj

Subscribers: thomie

Differential Revision: https://phabricator.haskell.org/D1772

GHC Trac Issues: #11416