Set stop_gradient=True for some variables in SSD API.

[qingqing01]

When the outputs of ParallelDo op are not used by any operators and does not set stop_gradient = True. The backward transpiler will fail. For example, without #9360 , the error in MobileNet-SSD with parallel_do_op is as follows:

  File "train.py", line 160, in <module>
    num_passes=300)
  File "train.py", line 86, in train
    optimizer.minimize(loss)
  File "/home/users/dangqingqing/.jumbo/lib/python2.7/site-packages/paddle/fluid/optimizer.py", line 235, in minimize
    [error_clip_callback])
  File "/home/users/dangqingqing/.jumbo/lib/python2.7/site-packages/paddle/fluid/backward.py", line 479, in append_backward
    _append_backward_vars_(root_block, fwd_op_num, grad_to_var, grad_info_map)
  File "/home/users/dangqingqing/.jumbo/lib/python2.7/site-packages/paddle/fluid/backward.py", line 367, in _append_backward_vars_
    _append_backward_vars_(sub_block, 0, grad_to_var, grad_info_map)
  File "/home/users/dangqingqing/.jumbo/lib/python2.7/site-packages/paddle/fluid/backward.py", line 382, in _append_backward_vars_
    op_desc.infer_shape(block.desc)
paddle.fluid.core.EnforceNotMet: Input(Out@GRAD) should not be null at [/home/users/dangqingqing/Paddle/paddle/fluid/operators/transpose_op.cc:107]
PaddlePaddle Call Stacks: