This big MR entirely removes the "flattener" that took a type and
replaced each type-family application with a fresh type variable.
The flattener had its origin in the paper
     Injective type families for Haskell

But (a) #25657 showed that flattening doesn't really work.
    (b) since we wrote the paper we have introduced the so-called
        "fine-grained" unifier GHC.Core.Unify, which can return
                 * SurelyApart
                 * Unifiable subst
                 * MaybeApart subst
        where the MaybeApart says that the two types are not unifiable by a
        substitution, but could (perhaps) be unified "later" after some type
        family reductions.  This turns out to subsume flattening.

This MR does a major refactor of GHC.Core.Unify to make it capable of
subsuming flattening.   The main payload is described in
       Note [Apartness and type families]
and its many wrinkles.

The key (non-refactoring) implementation change is to add `um_fam_env`
to the `UMState` in the unification monad.

Careful review with Richard revealed various bugs in the treament of
`kco`, the kind coercion carried around by the unifier, so that is
substantially fixed too: see Note [Kind coercions in Unify].

Compile-time performance is improved by 0.1% with a few improvements over
1% and one worsening by 1.3% namely T9872a.  (I have not investigated the
latter.)

Metric Decrease:
    T9872b
    T9872c
    TcPlugin_RewritePerf
Metric Increase:
    T9872a