convert module initialization to use multi-phase-init

[davidhewitt]

Part of #2274, inspired by all of #2245, #4081, and #4162

This converts the module initialization itself to use module slots and multi-phase init APIs. Module state is deliberately not covered here and left for follow ups. I also don't change anything related to subinterpreter checking, that is also left for a follow up.

[davidhewitt]

Note to self: this might need to be opt-out (for a few versions) and be mentioned in the migration guide.

[mejrs]

I like this very much!

Is this compatible with how we are currently putting pyclass' type objects in statics? We should figure out a way to store them in module state if possible.

[davidhewitt]

I think it's sorta fine to leave the type objects in statics for now given we don't support subinterpreters, but yes:

• we should seek to use module state if we can anyway as a followup
• and we'd need to do so for any potential eventual subinterpreter support

[alex]

I suppose eventually we'll want to allow users to have their own module state as well for their statics :-/

[mejrs]

I think it's sorta fine to leave the type objects in statics for now given we don't support subinterpreters, but yes:

* we should seek to use module state if we can anyway as a followup

* and we'd need to do so for any potential eventual subinterpreter support

This will also put the nail in the coffin for #1444. I'm fine with that personally, but it'll be disappointing for quite a lot of people. Maybe we should start thinking about other ways to address that use case.