Use more appropriate return type for resolve_associated_item

Previously, the types looked like this:

- None means this is not an associated item (but may be a variant field)
- Some(Err) means this is known to be an error. I think the only way that can happen is if it resolved and but you had your own anchor.
- Some(Ok(_, None)) was impossible.

Now, this returns a nested Option and does the error handling and
fiddling with the side channel in the caller. As a side-effect, it also
removes duplicate error handling.

This has one small change in behavior, which is that
`resolve_primitive_associated_item` now goes through `variant_field` if
it fails to resolve something.  This is not ideal, but since it will be
quickly rejected anyway, I think the performance hit is worth the
cleanup.

This also fixes a bug where struct fields would forget to set the side
channel, adds a test for the bug, and ignores `private_intra_doc_links`
in rustc_resolve (since it's always documented with
--document-private-items).

compiler/rustc_resolve/src/lib.rs

@@ -18,6 +18,7 @@
 #![feature(nll)]
 #![cfg_attr(bootstrap, feature(or_patterns))]
 #![recursion_limit = "256"]
+#![allow(rustdoc::private_intra_doc_links)]
 
 pub use rustc_hir::def::{Namespace, PerNS};