change Py constructors from pointer to produce Py<PyAny>

[davidhewitt]

The constructors such as Py::from_owned_ptr are implemented on Py<T>. This is different to Bound and Borrowed where the implementation is only on Bound<'py, PyAny> and Borrowed<'a, 'py, PyAny>.

I would argue that implementing only on the PyAny type is better because it forces the caller to decide what casting logic to do - e.g. .cast(), .cast_exact(), or .cast_unchecked(). The Py<T> constructors basically all do .cast_unchecked().

Implementing this diff I found places in the codebase where the unchecked cast was probably not good (return value of PyNumber_Index), and I'd bet user code more likely makes the same mistake.

If we like moving forward with this, I think a better path than actually changing the constructors immediately would be to introduce new temporary names (Py::any_from_owned_ptr etc. maybe?) and deprecate the existing, and rename the new forms once the deprecated ones are removed.

We might want to also introduce proper .cast() methods to Py<T> similar to Bound::cast etc, to make the handling of the Py<PyAny> return values easier.

[Icxolu]

In general I would be in favor of this. Maybe we can even go a step further and just deprecate them, and tell people to use the Bound or Borrowed constructors instead and unbind in the end to get the Py, which I would say ist best practice anyway. It's not really longer to write and all the functionality is already there.

[davidhewitt]

Hmm interesting, I hadn't really considered that possibility.

To be honest, temporary introduction of Py::any_from_owned_ptr etc. does seem pretty pointless if we'd know we'd want to rename it back to Py::from_owned_ptr eventually anyway, given we've got Bound and Borrowed constructors which are preferred as you say.

I'll come back to this for 0.28 I think 👍