Beyond Logit Lens: Contextual Embeddings for Robust Hallucination Detection & Grounding in VLMs