WIP: Type/data instances: require that the instantiation is determined by the LHS alone (#23515)

This is the new version of @sand-witch's !10672 (closed), addressing #23515.

The compiler change is in GHC.Tc.TyCl.

I've added fixes for the 61 test failures that occurred in the previous MR and rebased onto master.

I will open a head.hackage MR with the necessary patches once CI tells me which packages are failing. At this point I can at least say that it affects contravariant (which has a lot of reverse dependencies) and servant.

Quick note on what kind of failures were commonly seen in the test, in no particular order:

1. Need to explicitly write non-linear patterns on kinds

     type (==) :: k -> k -> Bool
     type family a == b where
   -   f a == g b = f == g && a == b
   +   f (a::k) == g (b::k) = f == g && a == b
       a == a = 'True
       _ == _ = 'False

2. User implicitly expects a kind to be Type

     type family N' a where
   -   N' (t a) = [a]
   +   N' (t (a :: Type)) = [a]
       N' a     = ()

3. Need to pattern match on return kind

     type family Foo a :: k
   - type instance Foo Int = Bool
   + type instance Foo @Type Int = Bool

4. Matching on (->) is, by default, levity polymorphic

     type family F a where
   -   F (a -> _) = Maybe a
   +   F ((a :: Type) -> _) = Maybe a