ENH: stats.ttest_1samp: add array-API support

[mdhaber]

Reference issue

gh-20544

What does this implement/fix?

This adds array-API support to scipy.stats.ttest_1samp.

Additional Information

import numpy as np
import numpy as xp
from scipy import stats
rng = np.random.default_rng(34598234592345)
x = xp.asarray(rng.random((1000, 1000)))
stats.ttest_1samp(x, xp.asarray(0.))

# main (NumPy)
# 6.58 ms ± 299 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
# pr (NumPy)
# 6.59 ms ± 135 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
# pr (PyTorch)
# 3.68 ms ± 172 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

[rgommers]

Nice! Would it make sense to post a basic benchmark in PR descriptions, to show things work with similar or improved speed compared to using NumPy? E.g., using CuPy if GPU is supported, and PyTorch CPU tensors if only other CPU implementations are supported?

[lucascolley]

took a look on my phone, may have missed something but looks pretty good!

[lucascolley]

I am without a GPU for a while now, and I think all of these PRs should run CuPy through the test suite before going in.

[mdhaber]

Like pearsonr, this one is for CPU only until we have a standard way to convert to NumPy arrays for internal calculations (#18286 (comment)).

[mdhaber]

Addressed the comments and posted timing comparison at the top: NumPy performance is unchanged but PyTorch is considerably faster. One of the Windows tests stopped like:

........................................................................ [ 87%]
..............s...................................................s..... [ 87%]
...................................................
Error: The operation was canceled.

last time. Not seeing this in other PRs so let's watch out for it.

[rgommers]

NumPy performance is unchanged but PyTorch is considerably faster.

Very nice. Even if PyTorch had not been faster this change has value - but almost 2x faster out of the box is great to see.

[lucascolley]

All LGTM. I'm confident that test coverage hasn't reduced and that there aren't any spooky stats changes that have gone over my head. Thanks Matt!