[Feature] Enable multi gpu distributed training of fluid

[typhoonzero]

Resolves #8139

Sample code to run multi GPU distributed training:

def train_loop_parallel(use_gpu, trainer_prog, trainer_id=0, bcast=False):
        place = core.CPUPlace() if not use_gpu else core.CUDAPlace(0)
        startup_exe = fluid.Executor(place)
        startup_exe.run(fluid.default_startup_program())
        exe = fluid.ParallelExecutor(use_gpu, avg_cost.name)

        feeder = fluid.DataFeeder(place=place, feed_list=[images, label])

        for pass_id in range(args.num_passes):
            for batch_id, data in enumerate(train_reader()):
                print("before run one...")
                loss, = exe.run(
                        [avg_cost.name],
                        feed_dict=feeder.feed(data))
                if bcast:
                    exe.bcast_params()
                print("Pass %d, batch %d, loss %s" % (pass_id, batch_id, np.array(loss)))