This reduces the peak decoding error by removing small (inaudible) spikes in
the error at the frame boundaries. These were due to the frequency-domain bias
ending up as a small pulse in the middle of the IMDCT overlap. None of this
was ever audible, but fixing it is still cleaner.