BUG: Fix delegation behaviour with atleast_3d

[Enderdead]

This PR fixes the bad behavior reported in scipy/scipy#23929 when ndim=3.

It removes delegation when ndim > 2 and adds additional tests.
These tests reproduce the issue while avoiding input arrays of size 1.

I'm not entirely happy with these new tests, but they provide a reasonable starting point for further iteration.

Feedback is welcome.

[lucascolley]

Thanks for the PR!

I wonder whether, rather than checking for exact equality in all of these tests, we should just check for expected shapes? Probably good to still keep one sanity-check test that we aren't unexpectedly modifying the values, but otherwise the tests are quite verbose.

Thoughts @Enderdead ?

[Enderdead]

@lucascolley You're right — it will be easier to understand and maintain. I'll propose another version ASAP.

[Enderdead]

@lucascolley I’m proposing this new version with a hash method to perform a semi–sanity check.
Sorry if it’s still a bit heavy — it’s difficult for me to gauge how far we can go with simplifying this test.

[lucascolley]

thanks @Enderdead, I like the approach!

[Enderdead]

I replaced the asarray + range combination with xp.arange. I also squashed the three commits.

Do you want me to add @pytest.mark.xfail_xp_backend(Backend.SPARSE, reason="no arange", strict=False) in order to pass the new tests ?

[lucascolley]

since sparse has no arange, I think we can generate arrays with np.arange and wrap with xp.asarray.

[Enderdead]

Done !

[lucascolley]

thanks again!

[lucascolley]

@allcontributors please add @Enderdead for bug

[allcontributors]

@lucascolley

I've put up a pull request to add @Enderdead! 🎉