nn.Graph call and launch impl

[chengtbf]

主要功能：

• nn.Graph.__call__ 调用 compile 和 launch
• nn.Graph.__launch__ 调用 RunLazyNNGraph
• nn.Graph.__compile__ 注册 input/output/var op names； 启动 compile 和 runtime
  - [ ] nn.Graph 最小可运行 单测： Relu graph 调试工作交给 @leaves-zwx 后续完成

修复大量 BUG：

• pybind11 传入 vector string 的 BUG
• RunLazyNNGraph 传参 TensorTuple 的 BUG
• MultiClientSessionContext 漏掉 Global ProfilerConf 的 BUG
• NNGraph 编译 Plan 时没有回复 OpAttribute 的 BUG （提供 PlanUtil::PopulateOpAttibute）
• MultiClientAddCallbackNotifier 时 lbn 设置错误的 BUG
• CheckOpGraph 没有处理 Multi-Client 情形的 BUG
• CallBackNotify 输入 dtype 的问题； tick dtype uint8 -> int8
• WaitAndSendIds Kernel 运行 SegmentFault 的 BUG
  - [ ] 启动起来以后会卡住（两个线程100%占用并忙等）

提供新 Feature：

• AutoRegistrationFactory 获取 creator 时的更多提示报错信息（模板类型 和 key类型）
• convert_to_tensor_tuple 支持 空的 list
• Kernel 的 注册机（宏）提供容错机制

[chengtbf]

卡住的原因 Debug：

日志：

PhysicalRun::RunLazyJob instruction launched.
InstructionsBuilder::RunLazyJob
InstructionsBuilder::RunLazyJob:: new instruction.
InstructionsBuilder::RunLazyJob:: emplace back to instruction list.
cclog: WaitAndSendIdsKernel::GetBuffer name = ReluGraph_0.
LazyJobStreamType::Compute instruction proto: instr_type_name: "RunLazyJob"
.
RunLazyJobInstructionType::Compute begin.
RunLazyJobInstructionType::WaitUntilQueueEmptyIfFrontNNGraphNotEquals.
RunLazyJobInstructionType::GetAllBuffer and Send job_instance.
RunLazyJobInstructionType::EnqueueNNGraph.
LazyJobStreamType::Compute instructio compute.
cclog: WaitAndSendIdsKernel::Recevice job_instance down.
cclog: WaitAndSendIdsKernel::GetBuffer name = ReluGraph_0.

初步分析：

Kernel

• CallbackNotifiler 没有触发；
• WaitAndSendIds 已经触发了，而且第二次的 kernel 已经在等待了；

指令（从外到内）

• InstructionBuilder::RunLazyJob 被完整执行了
• RunLazyJobInstructionType 也被完整执行了
• LazyJobStreamType 的 一次 Compute 也执行完了

[github-actions]

CI failed, removing label automerge

[github-actions]

CI failed, removing label automerge

[github-actions]

CI failed, removing label automerge

[github-actions]

Speed stats:

GPU Name: GeForce GTX 1080 

PyTorch resnet50 time: 143.1ms (= 7155.0ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 126.2ms (= 6310.9ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
Relative speed: 1.13 (= 143.1ms / 126.2ms)

PyTorch resnet50 time: 86.3ms (= 4313.1ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 74.3ms (= 3713.1ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
Relative speed: 1.16 (= 86.3ms / 74.3ms)

PyTorch resnet50 time: 62.6ms (= 3132.4ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 55.9ms (= 2796.4ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
Relative speed: 1.12 (= 62.6ms / 55.9ms)

PyTorch resnet50 time: 55.2ms (= 2759.7ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 48.1ms (= 2402.7ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
Relative speed: 1.15 (= 55.2ms / 48.1ms)

PyTorch resnet50 time: 50.3ms (= 2515.6ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 68.6ms (= 3430.3ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
Relative speed: 0.73 (= 50.3ms / 68.6ms)