Towards Lossless Token Pruning in Late-Interaction Retrieval Models