gh-127773: Disable attribute cache on incompatible MRO entries

[encukou]

This brings back the functionality of Py_TPFLAGS_HAVE_VERSION_TAG, which should have been named “use version tag”, as the field itself is always present -- see #27260 where I mistakenly removed it.

The type attribute cache relies on a class being able to invalidate caches of all of its subclasses. If a type's MRO contains an entry that's not that type's base, the cache needs to be disabled.
This requires a single bit, which was in tp_flags in 3.9 and below.

I'd like to avoid using tp_flags for an internal flag (even one that's currently reserved for ABI compatibility).

We already have a mechanism for disabling the cache: it's disabled after 1000 modifications to a single type, using tp_versions_used as a counter. It would be possible to set this to 1000 (MAX_VERSIONS_PER_CLASS) when an incompatible MRO is found. To make the code & logic more understandable, this PR uses a new bigger constant, _Py_ATTR_CACHE_UNUSED. The existing tp_versions_used check is reused.

If the tp_versions_used mechanism is dropped in the future, _Py_ATTR_CACHE_UNUSED can become a single bit stored elsewhere in PyTypeObject. Ideally in a private (underscore-prefixed) field; not in tp_flags.

@markshannon @colesbury Does this sound right? I'm not much of an attribute cache expert and I don't know about future plans in this area.

• Issue: Stale entries in MRO cache if MRO contains non-base classes #127773

[colesbury]

This makes sense to me, but @markshannon is probably more familiar with the MRO cache and tp_versions_used.

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @encukou for commit 5dbbdf4 🤖

If you want to schedule another build, you need to add the 🔨 test-with-buildbots label again.

[encukou]

The issues on failed buildbots are unrelated (the ctypes one on RHEL8 is fixed in #128034).

[encukou]

@markshannon, do you want to take a look?

[encukou]

If there are no objections, I'll merge this next week.