Separate type for cross-crate encoded attributes

[jdonszelmann]

As per a suggestion of @lcnr. I've got some ideas on how to do this, but they mostly make sense when all attributes become parsed. At that point I think we can get rid of AttributeKind, and AttributeExt, and instead make one enum (Attribute) where the unparsed tool attributes become just an (uncommon) variant.

Part of the attribute rework: #131229 (comment)

@rustbot labels +A-attributes
@rustbot claim