Better stats for LOAD_ATTR and STORE_ATTR

[markshannon]

This PR provides better stats for LOAD_ATTR and STORE_ATTR, differentiating several class attributes.
There is a slight change to specializing behavior, but it seems to make no difference in practice.

The main change is for simple attributes of a class.

Consider

classC: i=0c=C() c.i

c.i is not specialized and the failure is reported as "not in dict".
i isn't in the instance's dict, but that's not the problem.
With this PR, the failure is reported as "class attr simple" which is much more informative.

The change in specialization behavior is that if an instance attribute shadows a non-data descriptor, it is no longer specialized.
This has no effect on the hit rate, which remains at 82%, but does convert some misses to deferreds.

[markshannon]

https://github.com/faster-cpython/ideas/blob/main/stats/pystats-2022-12-16-faster-cpython-05a9952.md

[Fidget-Spinner]

Interesting, the change in specialisation behavior is that instead of falling through and trying to specialise dict assess, we just fail immediately right?

I'm surprised the spec failures only increased by 1%.

[markshannon]

I think what is happening is that shadowing is rare, and is usually code that stores a default value on the class, shadowing it if a non-default value is needed on the instance.

So we tend to get a lot of misses anyway.

It is generally considered good practice to set all attributes in the __init__ method, so this case should get rarer as time goes by.