dont double check vectors

more correct EPZS implementation
much faster halfpel ME
simplifications / reducing duplicate code
4MV in LQ mode

Originally committed as revision 641 to svn://svn.ffmpeg.org/ffmpeg/trunk