Revision 1212

Valgrind reported lot of unitialized reads. These unitialized reads
helped sysKin finding three bugs:
  - ZeroMacroblock did not reset the cbp field. So for some skipped
    blocks, a test was done on the cbp value...
  - MEanalysis was using wrong mvs from the current bvop (unitialized
    or just wrong in current context). That's because in devapi3,
    bframes used to share the same mvs array wheras now, it's one array
    a bvop.
  - Collocated skipped MBs for a bvop didn't reset mvs[0] and b_mvs[0].