Disable filter_mb_fast() as it optimized the incorrect code.
Fixes at least:
BA3_SVA_C.264
CABA3_SVA_B.264
CABACI3_Sony_B.jsv
CAFI1_SVA_C.264
camp_mot_frm0_full.26l
CAWP5_TOSHIBA_E.264
CVFI2_SVA_C.264
CVSE3_Sony_H.jsv
CVWP2_TOSHIBA_E.264
CVWP5_TOSHIBA_E.264
SL1_SVA_B.264

Originally committed as revision 14315 to svn://svn.ffmpeg.org/ffmpeg/trunk