fix bug of SyncOnMasterFn atexit

[chengtbf]

修复 oneflow/__init__.py 在 atexit 注册 _SyncOnMasterFn 函数时，并没有真正执行这个 sync() 函数的 BUG。

测试： （在 sync() 函数内加测试 log ）

master:

chengcheng@oneflow-21:~/debug/graph $ python3 test_graph_linear.py 
Nothing happened because environment has been initialized
..
----------------------------------------------------------------------
Ran 2 tests in 1.530s
OK

本分支:

chengcheng@oneflow-21:~/debug/graph $ python3 test_graph_linear.py 
Nothing happened because environment has been initialized
..
----------------------------------------------------------------------
Ran 2 tests in 1.561s

OK
cclog: sync on master

[github-actions]

Speed stats:

GPU Name: GeForce GTX 1080 

PyTorch resnet50 time: 138.7ms (= 6935.9ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 127.6ms (= 6380.7ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
Relative speed: 1.09 (= 138.7ms / 127.6ms)

PyTorch resnet50 time: 83.4ms (= 4168.3ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 74.1ms (= 3707.4ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
Relative speed: 1.12 (= 83.4ms / 74.1ms)

PyTorch resnet50 time: 57.5ms (= 2877.3ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 47.1ms (= 2352.9ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
Relative speed: 1.22 (= 57.5ms / 47.1ms)

PyTorch resnet50 time: 47.9ms (= 2393.9ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 39.1ms (= 1954.2ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
Relative speed: 1.23 (= 47.9ms / 39.1ms)

PyTorch resnet50 time: 42.4ms (= 2120.0ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 40.8ms (= 2040.1ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
Relative speed: 1.04 (= 42.4ms / 40.8ms)