Generate `Normalize` clauses for dyn and opaque types

[lowr]

When a trait Trait has a transitive supertrait defined like trait SuperTrait: AnotherTrait<Assoc = Ty>, we've been generating impl AnotherTrait for dyn Trait and opaque type T: Trait but without accounting for the associated type bound. This patch adds Normalize clauses that correspond to such bounds.

For example, given the following definition

trait SuperTrait where WC {
   type Assoc where AssocWC;
}
trait Trait: SuperTrait<Assoc = Ty> {}

we generate the following clauses

// for dyn Trait
Implemented(dyn Trait: SuperTrait) :- WC
Normalize(<dyn Trait as SuperTrait>::Assoc -> Ty) :- WC, AssocWC // this patch adds this
// for placeholder !T for opaque type T: Trait
Implemented(!T: SuperTrait) :- WC
Normalize(<!T as SuperTrait>::Assoc -> Ty) :- WC, AssocWC // this patch adds this

I think this doesn't contradict #203 as "we may legitimately assume that all things talking directly about ?Self are true", but I'm not really sure if this is the right direction. One caveat is that this patch exacerbates "wastefulness" as in this comment as it adds the Normalize clauses even when we're trying to prove other kinds of goals.

Fixes #777

[jackh726]

@bors r+

[bors]

:pushpin: Commit 1346728050928f5ed800dea034603af9d38bcb0d has been approved by jackh726

It is now in the queue for this repository.

[bors]

:hourglass: Testing commit 1346728050928f5ed800dea034603af9d38bcb0d with merge 1635ed5da8e1eb6aceb9d11b4ebcd2ff732d653b...

[bors]

:sunny: Test successful - checks-actions Approved by: jackh726 Pushing 1635ed5da8e1eb6aceb9d11b4ebcd2ff732d653b to master...