b=3986

b=3986
r=adilger

- we can't initialize few groups in parallel because
  filter_read_group_internal() reallocates bunch of
  global-per-obd structures. so, here is simple patch
  to serialize it. indeed, this isn's very scalable
  to hundreds of MDSs, but I think it's enough for a
  while. tested on mountain and my home cluster