Created by: qingqing01

Add CUDA profiler tools in new framework to analyze CUDA performance conveniently.

The usage of #5954 is as follows:

import paddle.v2.fluid.profiler as profiler
with profiler.cuda_profiler('cuda_profiler.txt', 'csv') as nvprof:
     # run CUDA program.

This fuctions is used to profile CUDA program by CUDA runtime application programming interface. The profiling result will be written into output_file with Key-Value pair (kvp) format or Comma separated values (csv) format. The output file is cuda_profiler.txt is this example. The users can config the counters/options for profiling by config argument in profiler.cuda_profiler. The default config is

['gpustarttimestamp', 
 'gpustarttimestamp', 
 'gridsize3d', 
 'threadblocksize', 
 'streamid', 
 'enableonstart 0', 
 'conckerneltrace'].

Note this tool is only used for CUDA program. And it only counts the CUDA kernels and CUDA API calls within with scope.