gh-106581: Start projecting through calls

[gvanrossum]

For now, builds on gh-107760.

• Add function-by-version cache (and redid it using borrowed refs)
• Double max trace length to 64
• Trace into function calls
• Support RESUME op in Tier 2

TODO:

• Project through RETURN_VALUE (see Support RETURN from tracing through a function #107925)
• Add back _Py_EnterRecursivePy (see gh-106581: Split CALL_PY_EXACT_ARGS into uops #107760) and _Py_LeaveRecursiveCallPy (see Support RETURN from tracing through a function #107925)

• Issue: Call design for Tier 2 (uops) interpreter #106581

[gvanrossum]

This leaks quite a bit, probably in the function-by-version cache.

[gvanrossum]

I suspect I have to redo the function version cache using weak references if I want all tests to pass. Unfortunately that will slow things down. Having borrowed references to function objects in a cache seems horrible. What else could we try?

Perhaps the cache could contain code objects (which are what we want anyways, so perhaps a scheme like the current one doesn't break so many tests), and when a function's __code__ or __defaults__ attributes (or a few others) are updated we just lose the ability to specialize calls to it. In fact, when its __code__ is set, we could just reset its version to the new code object's co_version.

(BTW, why is func_version set to code->co_version only by MAKE_FUNCTION and not during regular function object initialization? This feels like a relevant mystery.)

[gvanrossum]

Or perhaps we could leave borrowed references in the function version cache that we remove when the function is deleted? (Also when the function's func_version attribute is changed.) That would be a custom variant on using weak references that might be faster.