Avoid an empty trait name in impl blocks.

`resolve_ident_in_lexical_scope` checks for an empty name. Why is this
necessary? Because `parse_item_impl` can produce an `impl` block with an
empty trait name in some cases. This is pretty gross and very
non-obvious.

This commit avoids the use of the empty trait name. In one case the
trait name is instead pulled from `TyKind::ImplTrait`, which prevents
the output for `tests/ui/impl-trait/extra-impl-in-trait-impl.rs` from
changing. In the other case we just fail the parse and don't try to
recover. I think losing error recovery in this obscure case is worth
the code cleanup.

This change affects `tests/ui/parser/impl-parsing.rs`, which is split in
two, and the obsolete `..` syntax cases are removed (they are tested
elsewhere).

compiler/rustc_resolve/src/ident.rs

@@ -296,9 +296,6 @@ impl<'ra, 'tcx> Resolver<'ra, 'tcx> {
     ) -> Option<LexicalScopeBinding<'ra>> {
         assert!(ns == TypeNS || ns == ValueNS);
         let orig_ident = ident;
-        if ident.name == kw::Empty {
-            return Some(LexicalScopeBinding::Res(Res::Err));
-        }
         let (general_span, normalized_span) = if ident.name == kw::SelfUpper {
             // FIXME(jseyfried) improve `Self` hygiene
             let empty_span = ident.span.with_ctxt(SyntaxContext::root());