BUG: Fix experimental dtype slot numbers

[seberg]

Unfortunately, I forgot to move them around when introducing ensure_canonical,
but the code relies on the order for simplicitly currently...

Which doesn't matter, just means that the DType part of the experimental ufunc
API is completely unusable in 1.23 right now.

---

Pretty big ooops by me, since I really thought I could tell peple in my SciPy talk that the super minimal unitdtype example will work with NumPy 1.23... And now I need this (so it will work with main only).

[seberg]

Shows that I should move the very minimal prototype into numpy itself soon probably, but I wold prefer not to do it quite now (just due to timing really).

[mattip]

Maybe we could add a CI job to run the prototype against HEAD?

[seberg]

Yeah, but I need to dumb it down a bit to avoid any use of unyt, etc. (or create a similar prototype example).

[seberg]

Ah, sorry would not need it for that. Also, I just realized we have the ScaledFloat thingy. Probably, I just need to make that example use the experimental API (potentially with some trick?).

That is a bit of annoying, but it failing is pretty meaningless, if it wasn't for the fact that I would like to reference a working version on Friday morning :).

[mattip]

3.11-dev is failing on a typing error, @BvB93 any thoughts? That is not the fault of this PR.

[mattip]

Maybe it is a regression with the recently released 3.11b4?

[BvB93]

Maybe it is a regression with the recently released 3.11b4?

Seems like python/cpython#93754 is the culprit, I'll create a PR in a bit.

[BvB93]

Right, a PR for the 3.11b4 regression has been created: #21982

[mattip]

Thanks @seberg for both the PR and for providing a nice platform to discover beta-4 incompatibilities :)