Fix ddp grad size

[daquexian]

修复 ddp 梯度尺寸不对的问题，改进单测

[github-actions]

Speed stats:

GPU Name: GeForce GTX 1080 

PyTorch resnet50 time: 154.6ms (= 7728.7ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 133.4ms (= 6667.7ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
Relative speed: 1.16 (= 154.6ms / 133.4ms)

PyTorch resnet50 time: 98.6ms (= 4930.2ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 72.8ms (= 3642.0ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
Relative speed: 1.35 (= 98.6ms / 72.8ms)

PyTorch resnet50 time: 73.6ms (= 3680.7ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 79.2ms (= 3961.1ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
Relative speed: 0.93 (= 73.6ms / 79.2ms)

PyTorch resnet50 time: 103.0ms (= 5149.7ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 52.8ms (= 2638.8ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
Relative speed: 1.95 (= 103.0ms / 52.8ms)

PyTorch resnet50 time: 87.1ms (= 4353.0ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 66.1ms (= 3303.6ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
Relative speed: 1.32 (= 87.1ms / 66.1ms)