branchless renormalization (1% faster get_cabac) old branchless renormalization wasnt faster because gcc was scared of the shift variable (missusing bit variable now)

Originally committed as revision 6587 to svn://svn.ffmpeg.org/ffmpeg/trunk