Implement GoOp

[wangkuiyi]

Go has goroutines, and the keyword go starts a goroutine. Fluid needs GoOp as the go keyword in Go.

Goroutines are green threads, which are not scheduling preemptively; instead, they actively give up the CPU time by hanging themselves up in a waiting queue.

Our first version doesn't have to implement green threads and the scheduling system. We could rather just create OS threads using std::thread.

The challenging part of this work is to make the newly created thread runs a block in a ProgramDesc. The WhileOp could be a good reference.

[typhoonzero]

How about we make two thread pools, similar to the Cuda stream that, one for computation intensive threads and one for IO intensive threads. The computation thread pool has number of threads same to CPU cores; The IO thread pool can have dynamically increase number of thread with a maximum count.

[cs2be]

Here is the proposed API: #8434