Add trait for cloning while the GIL is held

Due to issues making delay reference count increments sound, they were removed #4095 and replaced by an Clone implementation for Py which will panic without the GIL being held, gated by the py-clone feature.

This has several usability downsides, e.g. #[pyo3(get)] stops working with Py<T> fields which however is a very important use case when sharing data with Python code. Similarly, the PyBacked* types cannot be unconditionally Clone themselves.

Both problems (and likely others) should be resolvable if we add a trait, e.g. CloneRef or PyClone with signature

fn clone_ref(&self, py: Python<'_>) -> Self;

and implement it unconditionally for Py<T> and add a blanket impl based on Clone. The proc macro machinery behind #[pyo3(get)] could then go via this trait instead of the plain Clone.

Start by reviewing issue #4095 and the current Clone implementation for Py, then inspect the proc macro machinery behind #[pyo3(get)]. Define how the proposed trait and blanket Clone implementation should work for Py and PyBacked* types, and confirm that #[pyo3(get)] can use it without requiring the GIL-free Clone behavior.