Created by: luotao1

Inspired by the demands for profiling framework overhead, this PR add:

• --profile_infershape for infershape elapsed time of each operator
• --profile_compute for compute elapsed time of each operator.

Examples:

• command: ./paddle/fluid/inference/tests/api/test_analyzer_pyramid_dnn --infer_model=third_party/inference_demo/pyramid_dnn/model/ --infer_data=third_party/inference_demo/pyramid_dnn/data.txt --gtest_filter=Analyzer_Pyramid_DNN.profile --repeat=10000 --zero_copy --warmup --profile --profile_infershape --profile_compute
• result:

Event                                        Calls       Total       Min.        Max.        Ave.        Ratio.
thread0::hash                                120000      794.965     0.005452    0.473659    0.00662471  0.192506
thread0::fused_embedding_seq_pool            160000      691.328     0.003215    0.926386    0.0043208   0.167409
thread0::hash_compute                        120000      588.484     0.003927    0.179345    0.00490404  0.142505
thread0::sequence_enumerate                  120000      394.58      0.002732    0.838274    0.00328816  0.0955499
thread0::fused_embedding_seq_pool_compute    160000      385.397     0.001561    0.88532     0.00240873  0.0933263
thread0::fc                                  20000       347.146     0.016292    0.34892     0.0173573   0.0840636
thread0::fc_compute                          20000       269.738     0.012725    0.267968    0.0134869   0.0653187
thread0::sequence_enumerate_compute          120000      172.885     0.001212    0.610675    0.00144071  0.0418652
thread0::sum                                 20000       161.158     0.007017    0.145441    0.0080579   0.0390254
thread0::softsign                            20000       91.8232     0.004247    0.054925    0.00459116  0.0222355
thread0::sum_infershape                      20000       69.8391     0.002788    0.027515    0.00349195  0.016912
thread0::cos_sim                             10000       43.9477     0.003986    0.098801    0.00439477  0.0106422
thread0::sum_compute                         20000       43.2604     0.001959    0.139936    0.00216302  0.0104758
thread0::softsign_compute                    20000       30.9593     0.001391    0.042987    0.00154797  0.00749699
thread0::cos_sim_compute                     10000       23.7063     0.00214     0.096713    0.00237063  0.00574063
thread0::softsign_infershape                 20000       20.3502     0.000926    0.021815    0.00101751  0.00492793

In this example, we could find the framework overhead is big. For sum op: infershape costs 69/161=42%, and compute only costs 43/161=26%

thread0::sum                                 20000       161.158     0.007017    0.145441    0.0080579   0.0390254
thread0::sum_infershape                      20000       69.8391     0.002788    0.027515    0.00349195  0.016912
thread0::sum_compute                         20000       43.2604     0.001959    0.139936    0.00216302  0.0104758