Make the first device share data with the global scope in parallel_do_op.

[qingqing01]

Fix #9386

There is no need to copy data to the first device. Just make the first device share data with the global scope, since they are on the same device.

[panyx0718]

Add some comments to explain the problem?

The global moving_mean and moving_variance is currently not correctly updated by the values calculated from sub_scopes (unlike trainable parameters). Perhaps ParallelExecutor has the similar problem to solve
@tonyyang-svail @reyoung

[qingqing01]

Add some comments to explain the problem?

In #9386, the moving mean/variance in BN are un-trainable parameters. The trainable parameters will update in backward and copy to the sub-scope in each mini-batch before the forward. Different from other trainable parameters, the moving means/variances will not updated in backward, the parallel_do_op still copy the initialized parameters in the global scope.

This fix makes the first device share parameter address with the global scope. When the moving mean/variance in the first device is updated, they will also be updated in the global scope.

But for BN, only save moving mean/variance in the first device. Maybe we can merge them between multi-GPUs and multi-machines in the future.