Less is more: Summarizing Patch Tokens for efficient Multi-Label Class-Incremental Learning