Extend current profiler for timeline and more features.

[panyx0718]

Timeline has been a go-place for TF developers when doing performance profiling.
It visualizes the multi-device executions as different time-series. Arrows can be generated
for cross device data transfer. Additional features such as memory allocation/deallocations
are also very useful.

Some examples:

This is the first PR for the timeline feature.

1. It collects cuda kernel execution stats with user-defined names.
2. It stores the stats into a proto for future analysis.

Near-term next steps:

1. Some more clean up and collect other cuda events such as memcpy.
2. Generate timeline visualization with the protobuf.

[chengduoZH]

2016 ==> 2018, the same below

[panyx0718]

Done

[qingqing01]

1. The API in /~https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/cuda_profiler.h also can be used to get timeline, but only for the CUDA kernels. So a new timeline tracer for any code is good.
2. What is the difference between CUPTI API and NVTX API? The NVTX API seems also can get timeline, and the timeline can be visualized by NVIDIA Visual Profiler tool. It seems the PyTorch uses NVTX and the code is more simply /~https://github.com/pytorch/pytorch/blob/master/torch/csrc/autograd/profiler.h

[panyx0718]

For 1.
That API seems to be used to profile the time of a code block and dumpy to file? But we want to profile the whole model execution and programmatically get the execution start/end of all activities. Hence, I think it doesn't give the coverage we expect. Also, it dumps the data directly to file, but we want it to return from API.

For 2.
When I google NVTX, it directs me to https://docs.nvidia.com/gameworks/..., maybe it's designed for something else but happened to work with cuda? Plus, we don't want to depend on it's nvidia visualization tool. We want to have customized CPU and memory data. We also want to add our multi-machine information to our visualization.
In contrast, CUPTI is always shipped with cuda library. quoted: "The CUDA Profiling Tools Interface (CUPTI) enables the creation of profiling and tracing tools that target CUDA applications"