Speculative decoding is lossless because the main model checks whether it agrees with what the drafter outputted