Stop dereferencing msg_send!'s first argument

[madsmtm]

The issue

In Objective-C, having a null pointer receiver is valid, but with msg_send! it would be converted into a reference, which is undefined behavior - yes, even though the reference is converted into a pointer immediately after, it is still undefined behaviour!

Additionally, there might be a mutable reference somewhere else in the program, which would now be aliased.

Minimal example

let ptr: *mut Object = core::ptr::null_mut();
let _: () = msg_send![ptr, someSelector];

How to fix this

This PR proposes changing msg_send! to no longer automatically dereference the receiver. This is a breaking change, but neither e.g. cocoa nor winit is affected by it - they don't use smart pointers like objc_id::Id.

objc-foundation will require a few tweaks (it uses objc_id::Id), and so will probably also some other crates, but I would say this is acceptable, the fix is very simple, see SSheldon/rust-objc-foundation#14 (and it causes a compile error, not just silent UB if you don't fix it).

Also, I think with unsafe things like msg_send! it's always better to be explicit.

[SSheldon]

I agree this isn't great. The ergonomics of using smart pointers really suffer without it, though. I was hoping I could use specialization to implement a special trait for all messageable pointer types. I haven't kept up with specialization to see if that's possible yet. I agree we should do something about this unsoundness in the next objc version with a breaking change.

[madsmtm]

I don't think the ergonomics suffer that bad, in my view this is actually better since makes it clear that msg_send isn't consuming the Id, but simply using the pointer from it.

But maybe there's a way to add something like unsafe fn as_object_pointer(&self) -> *mut Object; on Message, and then change Message to be implemented on *const Object, &Object and such (like Encode in objc-encode) instead of on Object directly?