based on the max absolute value in a range.

to power-of-two.

(still requires float ops though). Also, removed the per-band IDCT.

per-band IDCT.

a few decoding errors.

represented by the laplace encoder (would have a probability of zero due
to finite precision)

to make sure that stereo coupling is done at the band level. Previously
the stereo coupling was done all at once, but there were all kinds of
interactions with the prediction and folding.

simplifications to the energy encoding.

(hopefully) no other side effect