move q_*_matrix out of MpegEncContext (40k ->23k) dct_quantize() is even slightly faster now, dont ask my why ...

Originally committed as revision 2411 to svn://svn.ffmpeg.org/ffmpeg/trunk