Do nodes always download a new block entirely, or will it take data from its own mempool to save on bandwidth if possible?