Cache the reference frame associated with the current MB mode in each fragment.

This avoids repeatedly performing the MB mode -> reference frame translation,
 which is done many times per fragment during DC prediction.
In fact, by using a special OC_FRAME_NONE tag for uncoded fragments, one need
 not even check to see if a fragment is coded during DC prediction, simplifying
 it even more.
This saves 2% on 720p decode on a Cortex A8.

svn path=/trunk/theora/; revision=17457