EXPERIMENT: Use transitive closure of tyvar substitutions

[rowanG077]

First of all this definitely isn't mergeable as is. It fixes #42 but if this is actually the solution it should go into ghc-tcplugins-extra. I have some suspicion there is a very good reason why this isn't already done as it seems too simple to not always have been done like this.

Please advice @christiaanb @martijnbastiaan

[christiaanb]

A lot of this might not be needed any longer from ghc 9.4 onwards, since all constraints are unflattened then. Could try the test case on the 9.4.1 PR: #41

[rowanG077]

@christiaanb Unfortunately the test also fails with GHC 9.4.1 on the branch you linked.

This the relevant type checker trace fragment:

tcPluginSolve start ghc-typelits-knownnat
  given   = [[G] $d~_a3Ep {0}:: b_a3Ea[sk:1]
                                ~ b_a3Ea[sk:1] (CDictCan),
             [G] $dKnownNat_a3Eg {0}:: KnownNat a_a3E9[sk:1] (CDictCan),
             [G] co_a3Eo {1}:: (2 ^ a_a3E9[sk:1])
                               ghc-prim:GHC.Prim.~# b_a3Ea[sk:1] (CEqCan)]
  wanted  = [[W] $dKnownNat_a3Er {0}:: KnownNat
                                         b_a3Ea[sk:1] (CDictCan)]
kn_wanteds
  [([W] $dKnownNat_a3Er {0}:: KnownNat b_a3Ea[sk:1] (CDictCan),
    KnownNat,
    b_a3Ea[sk:1],
    Outputable 
    b_a3Ea[sk:1])]
given_map
  [([W] $dKnownNat_a3Er {0}:: KnownNat b_a3Ea[sk:1] (CDictCan),
    KnownNat,
    b_a3Ea[sk:1],
    Outputable 
    b_a3Ea[sk:1])]
tcPluginSolveResult ghc-typelits-knownnat
  solved = []
  bad    = []
  new    = []

It doesn't include a flattening skolem anymore compared to GHC 9.0.2 so it's a little better. But it would still require manual "unflattening" (if you can call it that, there are no longer any fsk).