An Illusion of Unlearning? Assessing Machine Unlearning Through Internal Representations