pyproto: remove inventory from implementation

[davidhewitt]

This is a refactoring for #[pyproto] which uses dtolnay specialization instead of inventory as the underlying implementation.

The dtolnay specialization is done for a type PyClassProtocols<T: PyClass>. There are traits PyObjectProtocolSlots<T> etc. which are what get specialized. The default implementation is PyObjectProtocolSlots<T> for &'_ PyClassProtocols<T>, and the specialized implementation is PyObjectProtocolSlots<T> for PyClassProtocols<T>.

The traits with slot getter methods have been replaced by free functions to implement the slots, which the pyproto macro code directly uses.

It works out as slightly fewer LOC and also because all the protocols are known at compile time, the compiler should be able to optimise pyclass initialization better.

At the moment we lost the type checking we had with TypedSlot. If that's a concern I think I can reintroduce a type check in the slot defs in the macro code.

Follow-ups:

• I think I can see a way to refactor the code further to implement Support all receiver types for all protocol methods #1206, but I'm not entirely sure it'll work out. So I stopped here for now which works nicely and I think is a nice simplification.
• I think that we can make inventory an optional feature where it's only needed when users want multiple #[pymethods] blocks for a single PyClass. That'd be a nice simplification to remove overheads.

[kngwyu]

Awesome, thank you so much!
Though I don't think the inventory overhead is large, it seems good for me to have a simplified implementation.

The traits with slot getter methods have been replaced by free functions to implement the slots, which the pyproto macro code directly uses.

Aren't these changes actually unrelated to removing inventory? I like the way it flattens the definitions, though.

At a glance, this PR looks almost OK for me.
I feel that we can make things simpler in pyo3-derive-backend/src/defs.rs, but am not sure and need to check it again.

[davidhewitt]

Aren't these changes actually unrelated to removing inventory?

Sort of - I wanted the new specialised traits to provide &'static [PyType_Slot] lists. I ran into various issues with const fn unless I flattened the getter methods away.

I feel that we can make things simpler in pyo3-derive-backend/src/defs.rs, but am not sure and need to check it again.

I think I can hard-code some simplifications about the paths in defs.rs - I'll push a commit to do so.

[davidhewitt]

I pushed a commit which puts the module path on the top-level Proto def rather than each slot def. This reduces repetition and LOC a bit.

As I think these defs might have to change again to implement #1206 in the way I imagine, I think it's not worth doing further refactoring right now.

[davidhewitt]

@kngwyu ok if I merge this today before the release? It technically "breaks" any code using the #[doc(hidden)] public implementation details of pyproto (though hopefully that's nobody), so it could be nicer to release as part of a major version.

[kngwyu]

Sure, I noted some additional suggestions.

[davidhewitt]

👍 will merge this shortly, then rebase the release and push it out!