Increasing the CELT bit allocation slightly for stereo hybrid

Fixes a "regression" from 1.2 to 1.3-beta that was especially noticeable
on a Korean speech sample. (it wasn't a real regression since 1.2 only
worked because it was using CELT)