Register ForeignCallback and Watcher in Multi-Client

[chengtbf]

• RegisterForeignCallbackOnlyOnce -> RegisterGlobalForeignCallback 不再限制 ForeignCallback 和 Watcher 只能注册一次
• 提供 Destroy ForeignCallback 和 Watcher 方法
• Multi-Client 启动时也注册 ForeignCallback 和 Watcher （nn.Graph train 所需）
• Single-Client 启动时先尝试 Destroy Multi-Client 注册的 Callback

[github-actions]

Speed stats:

GPU Name: GeForce GTX 1080 

PyTorch resnet50 time: 141.0ms (= 7048.6ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 126.0ms (= 6298.8ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
Relative speed: 1.12 (= 141.0ms / 126.0ms)

PyTorch resnet50 time: 85.9ms (= 4293.1ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 73.4ms (= 3671.9ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
Relative speed: 1.17 (= 85.9ms / 73.4ms)

PyTorch resnet50 time: 57.8ms (= 2890.7ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 49.7ms (= 2484.8ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
Relative speed: 1.16 (= 57.8ms / 49.7ms)

PyTorch resnet50 time: 46.2ms (= 2308.2ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 48.8ms (= 2441.0ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
Relative speed: 0.95 (= 46.2ms / 48.8ms)

PyTorch resnet50 time: 40.6ms (= 2027.9ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 50.0ms (= 2497.8ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
Relative speed: 0.81 (= 40.6ms / 50.0ms)