Both registered and likely voters are registered. But a fair number of people who are registered to vote don't exercise their right. So the best polls are those who screen them out, leaving only the likely voters. The trick is predicting who will vote based on their past behavior.
I think that Gallup, for example, asks if you a registered to vote. If you are, they ask a series of seven questions to establish whether you are likely to vote.
Registered voter polls do not correspond as well to final results as likely voter polls.
Naturally this Mason-Dixon stuff is more of a registered voter poll than a likely voter poll which is why the results are so far off from other polls.
Thanks for the helpful explanation.