MergeIntoTable should avoid ambiguous reference

When resolving the UpdateAction, which could reference attributes from both target and source tables, Spark should know clearly where the attribute comes from when there're conflicting attributes instead of picking up a random one.