The following attempt is not optimal, but because I couldn't achieve a first-move blunder, the average centipawn loss will not rise over 1980.
http://en.lichess.org/8e5IN4x7/white (1964/1974)
I am fairly convinced that #11's 1970/1980 is the best possible 'solution' under Lichess's current analysis system. (Good job Lightsss!) A 'proof' follows.
White can't blunder on the first move (from experimentation). On the second move, White blunders from 0 to -10 (from 0 and not -10 because no blunder has been made yet). On every move after that, White blunders from +10 to -10. Averaging over 100 moves, we get 1970 average centipawn loss. This is the highest possible for White.
Similarly, Black also can't blunder on the first move; on every move after that, Black blunders from -10 to +10. Averaging over 100 moves, we get 1980 average centipawn loss. This is the highest possible for Black.
However useless this thread was, it surely was interesting to think about. :)