Fix or prevent projections of associated types in impl headers

[nikomatsakis]

Right now we permit associated type projections in the impl trait-ref and self-type:

impl<T:Foo> Foo<T::Baz> for T

But this isn't fully supported. For example, coherence doesn't consider such types (#20624) and I think there will be other problems due to inadequate normalization. I'm not so worried about the bugs per se, but we should decide if we think we will support this for 1.0 or not. If not, we should either just prevent projections in those positions for now.

At first I thought this was really hard to support, but now I am not so sure. It might just be a matter of inserting normalizations into a few places.

[nikomatsakis]

Nominating and adding to alpha milestone.

[nikomatsakis]

I decided to just fix them (and did so, in PR #20757)