Make tuning tests for OpenMP independent of platforms

[roastduck]

For test_auto_fission_fuse.py::test_tune_fuse and test_auto_fission_fuse.py::test_tune_fission, I compare two candidate scheduling plans during the test run, and use the comparing result to check our tuner, instead of expecting a preset result. I did not apply this change to the CUDA test test_auto_fission_fuse.py::test_tune_with_cond yet, because it has 4 different types of candidate plans.

To make things work, I also restart OpenMP at the beginning of timing by omp_pause_resource_all. It works, but I am not sure how the running time of OpenMP is affected. If not restarting OpenMP, every timing runs all returned a low stddev, but time from different runs differed greatly. It worth noting that omp_pause_resource_all is only available since OpenMP 5.0, which makes use impossible to use it from PyTorch's OpenMP 4.0, required by #421.

With the above changes, I removed the OMP_PROC_BIND magic set by #266.