ENH: speed up array_namespace

[crusaderky]

Related:

• ENH: cache helper functions #308
• ENH: Speed up array_namespace through caching scipy/scipy#23057

Tested locally vs. jax==0.4.31.

Benchmark

>>> import importlib
>>> from array_api_compat import array_namespace
>>> for xp_name in ("numpy", "cupy", "dask.array", "torch", "jax.numpy", "sparse", "ndonnx"):
...     print(xp_name)
...     xp = importlib.import_module(xp_name)
...     a = xp.asarray(1)
...     %timeit array_namespace(a)

Before

numpy       1.31 μs ± 19.9 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
cupy        1.36 μs ± 5.16 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
dask.array  1.77 μs ± 33.6 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
torch       1.52 μs ± 4.8 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
jax.numpy   1.4 μs ± 19.8 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
sparse      1.75 μs ± 9.39 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
ndonnx      1.99 μs ± 10.9 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

After

numpy       673 ns ± 11.8 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
cupy        494 ns ± 8.45 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
dask.array  498 ns ± 3.87 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
torch       510 ns ± 5.91 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
jax.numpy   521 ns ± 3.06 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
sparse      766 ns ± 2.13 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
ndonnx      727 ns ± 9.62 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

Note that the extra slowness of numpy is due to jax-ml/jax#20620

[crusaderky]

Note: this test requires jax>=0.6.1 to pass with api_version=2024.12. I didn't see much of a point adding special handling for older versions.