Created by: qingqing01

Fix https://github.com/PaddlePaddle/Paddle/issues/7992

• Use CUDA_VISIBLE_DEVICES to change the visible GPU order for each processor before ctest command.

• Use ctest -I to run tests in parallel, please to refer https://cmake.org/cmake/help/v3.0/manual/ctest.1.html?highlight=ctest to learn more about -I option

-I [Start,End,Stride,test#,test#|Test file], --tests-information Run a specific number of tests by number.

This option causes ctest to run tests starting at number Start, ending at number End, and incrementing by Stride. Any additional numbers after Stride are considered individual test numbers. Start, End,or stride can be empty. Optionally a file can be given that contains the same syntax as the command line.

But some output log is disorderly:

      Start 165: test_split_and_merge_lod_tensor_op
52/75 Test #206: test_is_empty_op ..................   Passed    9.42 sec
      Start 210: test_tensor
53/75 Test #211: 42/76 Test #165: test_crop_op ................................test_split_and_merge_lod_tensor_op .......   Passed     Passed   10.32 sec
  8.57 sec
      Start 215: test_cond_op
      Start 169: test_lookup_table_op
54/75 Test #215: test_cond_op ................................   Passed    3.12 sec