Move load_npy to C++

[mthrok]

So that the GIL is released entirely

Performance (based on data_formats.py)

• QPS of loading NPY files with multi-threading

  Concurrency	Baseline	C++	Improvement
  32	1577	1891	17%
  16	1561	1717	9%
  8	1592	1859	14%
  4	1590	1812	12%
  2	1769	1820	3%
  1	1690	1948	13%

• QPS of loading NPZ (no compression) files with multi-threading

  Concurrency	Baseline	C++	Improvement
  32	1577	1726	9%
  16	1495	1755	15%
  8	1607	1754	8%
  4	1591	1677	5%
  2	1637	1781	8%
  1	1658	1829	9%

• QPS of loading NPZ (with compression) files with multi-threading

  Concurrency	Baseline	C++	Improvement
  32	1192	1473	19%
  16	1241	1633	24%
  8	1230	1640	25%
  4	1305	1548	16%
  2	1379	1547	11%
  1	1277	1571	19%

The performance of multiprocessing stays roughly same.

[facebook-github-bot]

@facebook-github-bot has imported this pull request. If you are a Meta employee, you can view this in D78947559. (Because this pull request was imported automatically, there will not be any future comments.)