Refine argument naming for ParallelExecutor

[pkuyym]

Seems customize_loss_grad is not very self-explanatory and there's is no doc for this argument. I propose to change it to a more clear name like use_default_grad_scale and the default value is true. @reyoung @panyx0718 Need your opinion. Thanks.

Paddle/python/paddle/fluid/parallel_executor.py

Line 33 in bfafcbe

customize_loss_grad=False):

[panyx0718]

sounds good, need some comments to explain.