* using MOVQ_WONE/MOVQ_BFE as two instruction instead of static memory value access
  as its always faster
* PAVGB_MMX macro is using now mm6 -> mm7 is unmodified
* replaced original pixels_xy2_mmx with new faster and equal implementation
* replaced usage of mm7 for other then ZERO contstant in _rnd & _avg file with mm6


git-svn-id: file:///var/local/repositories/ffmpeg/trunk@632 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b