Truth is, it really doesn't matter. You can prove this to yourself by exporting several versions of the song using different dither algorithms, including one with no dither at all. 99% of the time, you will not be able to distinguish one from another in a blind A/B test. Normally you'd have to set up an unrealistic scenario (e.g. a greatly boosted reverb tail) to hear the difference.
But to answer the question, I use Ozone's MBIT+ algorithm. Not because it's noticeably better, but because it's more convenient to specify it there in the limiter than to remember to choose one at export time. Even if you don't have Ozone, any other third-party limiter is likely to provide a dithering option.