libavcodec/mpegvideo_enc: fix multi-threaded motion estimation rounding for mpeg4

ff_init_me() was being called after ff_update_duplicate_context(),
which caused the propagation of the initialization to other thread
contexts to be delayed by one frame.

In the case of mpeg4 (or flipflop_rounding), this would make the
hpel_put functions differ between the first thread (which would be
correctly initialized) and the other threads (which would be stale
from the previous frame).

libavcodec/mpegvideo_enc.c

@@ -3623,6 +3623,9 @@ static int encode_picture(MpegEncContext *s)
         s->q_chroma_intra_matrix16 = s->q_intra_matrix16;
     }
 
+    if(ff_init_me(s)<0)
+        return -1;
+
     s->mb_intra=0; //for the rate distortion & bit compare functions
     for(i=1; i<context_count; i++){
         ret = ff_update_duplicate_context(s->thread_context[i], s);
@@ -3630,9 +3633,6 @@ static int encode_picture(MpegEncContext *s)
             return ret;
     }
 
-    if(ff_init_me(s)<0)
-        return -1;
-
     /* Estimate motion for every MB */
     if(s->pict_type != AV_PICTURE_TYPE_I){
         s->lambda  = (s->lambda  * s->me_penalty_compensation + 128) >> 8;