gh-122029: Do not unpack method for legacy tracing anymore

[gaogaotiantian]

INSTRUMENTED_CALL and INSTRUMENTED_CALL_KW now unpack the method before monitoring, so they are not affected by this change. The bytecode that this affects is INSTRUMENTED_CALL_FUNCTION_EX.

For INSTRUMENTED_CALL_FUNCTION_EX, we do not unpack the callable, instead, we use PyObject_Call directly on the callable. Because of this, sys.monitoring does not know this eventually calls into a C function, because the callable is a Python method, so c_return event will not be generated.

With the current code, because we unpack the method in sys.setprofile, we will generate an unmatched c_call, which is horrible.

The ideal result is probably have a consistent result for all three bytecodes, but that requires some refactoring for CALL_FUNCTION_EX which @markshannon was kind of against. If we can't achieve that, we should at least generated paired events - so we don't have a c_call without its c_return. That's why I removed the unpack code for legacy tracing.

• Issue: CPython profiler broken with TensorFlow 2.17.0 code in Python 3.12.1+ #122029

[markshannon]

I took another look at CALL_FUNCTION_EX and the refactor to unpack the bound method would require a fair amount of extra logic to handle the self argument, either unpacking self and the remaining args into a new array, or prepending self to the args tuple.

It might be worth it if it were to unlock other optimizations, but let's go with this much simpler fix for now.

[markshannon]

I've suggested one slight tweak to the test, otherwise looks good.