This doesn't bear any resemblance to current admissions policies. Discrepancies between races in average SAT score at a given college are in the hundreds of points, not the low tens.
Without making a judgment one way or the other on the core issue - given that this paper is looking at "highly selective private research universities" with average SAT scores of over 1300, a 50 point difference is not very big and probably within each individual student's margin of error. When you are in that scoring range, each incorrect answer has a magnified score penalty to maintain the normal distribution.
1. Asians suffer a penalty of 50 points relative to whites. We can equally say, using the numbers from rjdagost's comment, that whites themselves suffer a penalty of 230 points.
2. The entire concept of the normally curved scores is that one 50-point difference is just as significant as any other 50-point difference. That's why we use standard deviations to compare things. It is true that scores at the very high end are more volatile, since most of the test's "space" is devoted to distinguishing between people in the middle of the range, but 1300 does not seem to be at a level where that worry would make sense; the data I have to hand is a score report from 2000 saying "on average, a person with a math score of 710 loses 2 points on a second testing", which is pretty reliable. And regardless, while an individual student might easily lose 30 points by retesting, group averages don't move so easily.
On thinking about this further, I've grossly misused the word "reliable". What that quoted bit says is precisely that while an individual's score may vary from test to retest (and is thus not necessarily reliable), the average score of a group is highly stable.
SD for verbal is 115, math is 120. This must have been what I was thinking of. SD for overall (V+M) score is 218. The overall SD is much more relevant here, but let's note that a difference of one standard deviation is still a very large difference.