There's a footnote that says: For definition of median, see Guide to Tabular Presentation, which appears to be this (from a previous year): http://www.census.gov/prod/2007pubs/08abstract/guide.pdf [PDF]
And that guide defines "median" in the normal way you would expect: arrange values in order, and choose the middle value, or average the middle two values if there is an even number.
A guess, given the other data in the table, is that they first aggregate by number of sexual partners (shown in the columns), then look for a point where the number on the left and right would be equal. I believe for non-continuous data like this you can compute a "fractional median" that takes into account how far into one of the cohorts the midpoint is. The intuition is that you'd want the median to be higher if the midpoint was "just past 2" (i.e. barely including any 3s) vs. "just below 4" (i.e. including almost all the 3s), even if the midpoint itself is a 3 in both cases.
Black males are such studs (8.3 partners) and latinas really save themselves for marriage (1.7!).
Guess what, people lie.
You can get a similar result even with means, if for some reason your sampling procedure doesn't catch the outliers. (For instance, if they're very rare, or if they're demographically different in a way that makes it harder for your sampling to find them.)
Having said all that, I suspect lying really is the main contributor to these differences.
Huh? How exactly do you end up with .3 of a sexual partner?