PatternSynonyms on data family and COMPLETE signature lead to incorrect warning

Note that the original test case below is fixed in !1903 (closed). There's a more obscure test case in #17207 (comment 227180), that's why this ticket is still open.

---

Here's a whitebox test I came up with that shouldn't produce any warnings:

{-# OPTIONS_GHC -Wincomplete-patterns -Wno-missing-methods -fforce-recomp #-}
{-# LANGUAGE GADTs, TypeFamilies, PatternSynonyms #-}
module Lib where

data family T a

data instance T () where
  A :: T ()
  B :: T ()

pattern C :: T ()
pattern C = B
{-# COMPLETE A, C #-}

g :: T () -> ()
g A = ()
g C = ()

h :: T () -> ()
h C = ()
h A = ()

Yet it currently (after !963 (closed)) produces the following warning:

testcases/LateInitOfShapes.hs:16:1: warning: [-Wincomplete-patterns]
    Pattern match(es) are non-exhaustive
    In an equation for ‘g’: Patterns not matched: B
   |
16 | g A = ()
   | ^^^^^^^^...

This is loosely related to the observation in #14059 (comment 140372): The first match of g commits the type of the value abstraction to the representation TyCon of T (), at which point we don't see the COMPLETE set that is attached to the family T. To my knowledge there is no way to get back to T () from the representation TyCon (Edit: yes, there is. tyConFamInst_maybe, which !1903 (closed) does), so the only way to recover this would be to make sense of the CoPat surrounding the match on A. There was a ticket where we discussed how to adding a CoPat variant to PmPat, but I couldn't find it. It's #14899 (closed), but I no longer think the CoPat idea is relevant.