Make substs(), input_types(), etc on PolyTraitRef preserve the binding level

[nikomatsakis]

These accessors allow you to skip past the binder in a very lightweight way. This is a bad thing, since the binder is significant. We should refactor them away. While we're at it, it'd might be best if the internal field of a Binder was private. We would then add accesses for every PolyFooBar type for the fields of FooBar. These accessors would yield a Binder<T> for those fields that may contain late-bound regions. I've been moving to this model gradually, but we should finish the job.

[jroesch]

cc me, 👍

[steveklabnik]

Traige: FIXMEs still in the code, no change.

[frewsxcv]

Triage: FIXMEs still in the code, no change.