Notes on the PR for reviewers:

• Since this API is intended for use by extensions, and _Py* APIs are internal-only, I added the API as public (Py*). But I think it should perhaps be CPython-specific, so I put it in Include/cpython/dictobject.h, not Include/dictobject.h. Not sure about any of this, though, so happy to change as requested.
• The STORE_ATTR_WITH_HINT opcode implementation in ceval.c pokes directly at dict internals. This meant I had to duplicate a bit of code that otherwise would be private to dictobject.c. I think the best option here would be to add higher-level (if private) API in dictobject (something like _PyDict_SetWithHint) for the needs of STORE_ATTR_WITH_HINT so it doesn't have to poke directly at dict internals anymore (though that may be difficult due to the needs of deopt), but that seemed a bit far afield for this PR, so for now I just added the duplication with a comment.
• Having dict_notify_event return the new dict version reduces the overhead in the unwatched case to a single testb/jne (checking only once if the dict is watched.)
• Pyperformance data in https://gist.github.com/carljm/987a7032ed851a5fe145524128bdb67a

https://bugs.python.org/issue46896