Add OpenMP support

[mthrok]

Add OpenMP support to build process, so that lfilter is parallelized. (#1557)

cc @yoyololicon

TODO:

• Benchmark lfilter CPU
  • Plot
• Add step to verify that the OpenMP version is same as PyTorch.

  >>> print(torch.__config__.parallel_info())
  ATen/Parallel:
      at::get_num_threads() : 8
      at::get_num_interop_threads() : 8
  OpenMP 201511 (a.k.a. OpenMP 4.5)
      omp_get_max_threads() : 8
  Intel(R) Math Kernel Library Version 2020.0.2 Product Build 20200624 for Intel(R) 64 architecture applications
      mkl_get_max_threads() : 8
  Intel(R) MKL-DNN v2.2.3 (Git Hash 7336ca9f055cf1bfa13efb658fe15dc9b41f0740)
  std::thread::hardware_concurrency() : 16
  Environment variables:
      OMP_NUM_THREADS : [not set]
      MKL_NUM_THREADS : [not set]
  ATen parallel backend: OpenMP
  

• (after merge) Check that nightly build libtorchaudio points the same OpenMP library as PyTorch.

Benchmark

Script is from #1441 (comment)

$ USE_OPENMP=0 python setup.py clean develop && python benchmark.py

[-------------- IIR filter --------------]
                   |  forward  |  backward
1 threads: -------------------------------
      [32, 256]    |    275.2  |    696.4
      [32, 1024]   |    651.2  |   1332.2
      [32, 4096]   |   1949.0  |   4122.1
      [64, 256]    |    418.9  |    934.1
      [64, 1024]   |   1135.0  |   2210.8
      [64, 4096]   |   3979.2  |   7947.9
      [128, 256]   |    658.7  |   1335.7
      [128, 1024]  |   1952.1  |   3933.4
      [128, 4096]  |  25779.2  |  35013.4
2 threads: -------------------------------
      [32, 256]    |    255.0  |    700.7
      [32, 1024]   |    559.6  |   1316.8
      [32, 4096]   |   1524.7  |   3221.2
      [64, 256]    |    361.4  |    918.5
      [64, 1024]   |    914.5  |   1813.1
      [64, 4096]   |   2854.4  |   6060.5
      [128, 256]   |    563.4  |   1297.8
      [128, 1024]  |   1529.0  |   3047.7
      [128, 4096]  |  15970.5  |  22613.1
4 threads: -------------------------------
      [32, 256]    |    237.8  |    696.3
      [32, 1024]   |    500.5  |   1281.7
      [32, 4096]   |   1307.5  |   2879.6
      [64, 256]    |    329.2  |    888.1
      [64, 1024]   |    815.5  |   1784.2
      [64, 4096]   |   2286.0  |   5158.9
      [128, 256]   |    502.5  |   1262.4
      [128, 1024]  |   1304.5  |   2707.9
      [128, 4096]  |  10360.7  |  15790.5

$ USE_OPENMP=1 python setup.py clean develop && python benchmark.py

[-------------- IIR filter --------------]
                   |  forward  |  backward
1 threads: -------------------------------
      [32, 256]    |    279.8  |    703.5
      [32, 1024]   |    652.5  |   1324.0
      [32, 4096]   |   1945.7  |   3922.1
      [64, 256]    |    417.6  |    938.1
      [64, 1024]   |   1133.3  |   2203.4
      [64, 4096]   |   3819.4  |   7863.2
      [128, 256]   |    681.2  |   1368.6
      [128, 1024]  |   1967.8  |   3962.5
      [128, 4096]  |  27461.1  |  35519.0
2 threads: -------------------------------
      [32, 256]    |    229.9  |    644.8
      [32, 1024]   |    444.4  |   1102.4
      [32, 4096]   |   1164.6  |   2434.2
      [64, 256]    |    305.3  |    814.8
      [64, 1024]   |    687.1  |   1509.4
      [64, 4096]   |   2117.4  |   4561.1
      [128, 256]   |    458.7  |   1094.7
      [128, 1024]  |   1160.6  |   2314.5
      [128, 4096]  |  14750.9  |  19707.2
4 threads: -------------------------------
      [32, 256]    |    196.9  |    608.9
      [32, 1024]   |    329.5  |    949.3
      [32, 4096]   |    674.4  |   1642.6
      [64, 256]    |    243.3  |    733.9
      [64, 1024]   |    454.2  |   1207.1
      [64, 4096]   |   1191.5  |   2841.3
      [128, 256]   |    333.4  |    937.6
      [128, 1024]  |    685.7  |   1570.5
      [128, 4096]  |   8411.4  |  11559.6

[mthrok]

TODO: double check this the later condition is necessary. It might work without it in macOS. (but not sure if that's okay. Consult someone from PyTorch core)