Update array API tests without adding complex number support

[cbourjau]

This PR updates the array API test suite to the latest upstream version plus a minor custom fix that is in the process of being upstreamed (upstream merged the fix). The fix allows us to run the test suite despite currently not having support for complex numbers, yet (see #32 for more information on adding those).

The skips.txt file contains failures that cause segfaults in the onnxruntime. xfails.txt contains test names which fail for other reasons. We should triage both of these files.

Edit: Upated the description to reflect the latest circumstance of this PR

[adityagoel4512]

Does this PR supersede #33?

[adityagoel4512]

#39 will address the 1 failure. The ONNX standard doesn't support pow on int8 or int16.

[cbourjau]

Does this PR supersede #33?

Yes! Sorry for the confusion!

[cbourjau]

Brief update:

The latest test suite runs quite fast, allowing us to run on many more hypothesis examples than previously (100 instead of 2 at the time of writing). This did surface some further issues. In particular, I discovered that we can trigger segfaults in onnx from the test_roll (example) and test_broadcast_arrays (example) tests.

[adityagoel4512]

Brief update:

The latest test suite runs quite fast, allowing us to run on many more hypothesis examples than previously (100 instead of 2 at the time of writing). This did surface some further issues. In particular, I discovered that we can trigger segfaults in onnx from the test_roll (example) and test_broadcast_arrays (example) tests.

Nice, thanks! If it's faster now it would be great to have skips.txt exclusively contain any tests that segfault upstream rather than assertion fails.

[cbourjau]

The full list of failures as of this writing can be found in the CI of this commit.
As suggested, the failures that may cause segfaults are in skips.txt; all others are in xfails.txt. Nonetheless, I didn't configure the test suite to hard-fail on XPasses (i.e. xfails that pass unexpectedly) since some of them are flakey even with the higher number of examples.

[adityagoel4512]

There are a few XPASSes.

[cbourjau]

There are a few XPASSes.

I'm afraid some of them are just flaky. I generated these files from observed failures.

[cbourjau]

I am now running the array-api-tests using a fixed seed. Thus, we should no longer get flaky XPASSes. I think this is now ready to merge even if we xfail a lot of tests liberally. We should address these in follow up PRs, IMHO.

[cbourjau]

Apparently the tests are still a bit flaky :( The XFAILS you just added were passing previously: https://github.com/Quantco/ndonnx/actions/runs/10677594899/job/29593026155#step:5:1263

[adityagoel4512]

Apparently the tests are still a bit flaky :( The XFAILS you just added were passing previously: https://github.com/Quantco/ndonnx/actions/runs/10677594899/job/29593026155#step:5:1263

I'm not sure how much marginal value there is in 100 iterations of each test compared to around 20. I'd rather make it 20 and have super snappy CI again.

[adityagoel4512]

I reduced the number of repetitions to 16 in the interest of having sub 10 minute CI. I think we can increase it once we've pruned down the list to catch more obscure edge cases but there is a reasonable amount to do as is. The added xfails are a little concerning but we can come back to them in a future PR.