hidden fetch output time cost into executor run ops.

[dzhwinter]

Now our implement separates the fetch variable and executor run a ProgramDesc.

Paddle/python/paddle/v2/fluid/executor.py

Lines 177 to 182 in 1575c2c

	self.executor.run(program.desc, scope, 0, True, True)
	outs = [
	core.get_fetch_variable(scope, fetch_var_name, i)
	for i in xrange(len(fetch_list))
	]

We can implement a queue and copy the output to the queue, only when the fetch call is ready, fetch out the value from the queue. It will not block the executor run the next epoch of training, which will reduce the wait time cost.