add keep_kids flag for executor

[jacquesqiao]

No description provided.

[reyoung]

The kids should not be kept.

• If users add a FetchOp, the device context will be synchronized. Every operator should be completed after Run.
• If users do not add a FetchOp, we should wait for the device context instead of keeping children scope. (We may add code to sync device for this situation.)

[emailweixu]

@reyoung In while_op.cc, the the step scopes created in the scope passed to executor.run needs to be kept so that they can be accessed by while_grad_op. If kids are dropped, how to achieve this?
/~https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/while_op.cc#L63

[reyoung]

In control operators, the create_local_scopes will be false. So the executor in WhileOp will not destroy any scopes. The scope will be destroyed after the whole program has been executed.

[reyoung]

In the original design of Executor, if create_local_scopes=False, the local scopes will not be destroyed.

Paddle/paddle/framework/executor.cc

Lines 130 to 132 in 6332b82

	if (create_local_scope) {
	scope->DeleteScope(local_scope);
	}

However, it seems the create_vars flag change this behaviour on our develop branch.

Paddle/paddle/fluid/framework/executor.cc

Lines 347 to 352 in 3f8d9b0

	if (create_vars && create_local_scope) {
	scope->DeleteScope(local_scope);
	} else {
	// Delete the local scopes created in operators.
	scope->DropKids();
	}

[emailweixu]

This needs to be fixed. Otherwise nested while_op does not work. What should be the right way to fix it?

[reyoung]

Merge this HotFix first. I will write a issue to record this problem

[reyoung]

The problem is recorded in this issue #11597