Fix the use of Object::_instance_bindings when GDExtensions are uninitialized

[adamscott]

Fixes #101465

[dsnopek]

Thanks!

As I mentioned over on #101465, this is a bug that should have already been fixed in godot-cpp.

However, if that bug has come back, I don't think this is the right solution.

Either (1) bindings need to make sure to clean up their singleton bindings when unloading (which is what godotengine/godot-cpp#1458 aimed to do in godot-cpp) or (2) we should check the GDExtensionManager to see if the extension is still loaded before trying to run the callbacks.

[adamscott]

However, if that bug has come back, I don't think this is the right solution.

Either (1) bindings need to make sure to clean up their singleton bindings when unloading (which is what godotengine/godot-cpp#1458 aimed to do in godot-cpp) or (2) we should check the GDExtensionManager to see if the extension is still loaded before trying to run the callbacks.

I don't understand.

1. When the original developer doesn't does it's job correctly, it shouldn't necessarily make Godot crash on exit.
2. How can we check GDExtensionManager when it's already uninitialized? My PR is all about that situation, when GDExtensionManager is already unloaded.

[dsnopek]

When the original developer doesn't does it's job correctly, it shouldn't necessarily make Godot crash on exit.

It isn't the job of the developer, but the language bindings, so godot-cpp in this case. There's lots of cases where if the language bindings don't do what they are supposed to, it will lead to crashes, and I don't think there is much we can do about that.

How can we check GDExtensionManager when it's already uninitialized? My PR is all about that situation, when GDExtensionManager is already unloaded.

It's a little tricky. I wrote a bit about it here:

2. We already associate the callbacks with a void *p_token which is the GDExtension * in the case of GDExtensions, which we could check for in GDExtensionManager. The only problem is that C# also uses these callbacks, and, in that case, the void *p_token is something else entirely. This won't be a problem once C# uses GDExtension, but for now, we'd probably need to track some new piece of data with the binding callbacks? Or, maybe, we can maintain a list of void *p_tokens that are no longer valid that we update when a GDExtension is unloaded, and then we don't run the free_callback if it's on the list?