WIP: New design

[adamgundry]

Fixes #98.

[adamgundry]

This mostly builds on 9.14.1 after surprisingly little work relative to 9.4! Next steps for this PR:

• Drop support for older GHCs in order to simplify the code.
• Use ghc-tcplugin-api more consistently rather than importing from ad hoc places in ghc.
• Look at the remaining issue in the tests: it seems mismatch3 generates an error message, but the exception doesn't get hit under -fdefer-type-errors. Avoiding (~~) seems to solve this.
• Update doctest and get the doctests working again (Doctest tutorial on 9.0 and 9.4 #96).

Other thoughts:

• We really should address Probable type soundness bug #97 by adding coercion evidence dependencies.
• Does this new design make Avoid non-deterministic type comparison #81 harder? We need to be able to sort types without introducing nondeterminism.
• Do we gain anything from (~~) or could we drop it to simplify the code?
• We can now define type family rewrites more directly in the GHC API, e.g. Pack (Unpack u) becomes u. Can we make more use of this e.g. to immediately simplify One *: u to u? Can we test this directly e.g. using :kind! in doctests?