refactor pyclass initialization

[Icxolu]

This refactors the code generation for #[new] / pyclass initialization. It uses the same const probe trick we already use elsewhere in the macro machinery to dispatch the initialization based on the return type of the #[new] function. This removes the Existing case from the PyClassInitializer, making it easier to solve passing arguments to the base __new__ in a followup. Additionally it allows to return arbitrary object implementing IntoPyObject from __new__ which also takes over the Existing case.

Closes #3291
Closes #4452
Closes #5560

[davidhewitt]

Very nice, thank you!

[Icxolu]

Maybe a side question: The tuple initializer could be a bit confusing now that we allow arbitrary object (I also always forget which order Subclass and Baseclass should have). Should we consider deprecating that syntax and move people to use PyClassInitializer directly?

[davidhewitt]

Looks great, thanks!

Should we consider deprecating that syntax and move people to use PyClassInitializer directly?

I fully agree that syntax is painful, let's do it in a follow up just because it's a big enough change worth landing, reviewing & documenting in its own PR.

[davidhewitt]

(we might also want to consider how to change the PyClassInitializer API to pass arguments to super __new__, or replace that API even)