Wow nice find, a quick skim brings up this line:
"analogRead() in a tight loop takes 39uS per loop" ...Some config changes here ... "And the ADC loop now takes 4.4uS. Cool" 
I'll do a bit more digging into this to find out if there are any undesirable results from making this change, it seems too good to be true although the author seems pretty aware of what he's doing. If everything does work out nice and stable this will massively reduce the overhead of getting samples for our fft's.