[Jahongir-Qurbonov]

CPy_AddTraceback creates a new code object for every traceback entry added by native code. On free-threaded builds, code objects use deferred reference counting, so only the cyclic GC can free them. The GC only runs when the eval breaker is handled, which never happens in a native loop that raises and catches exceptions without calling interpreted code. As a result, each iteration leaks a dead code object until the native function returns.

For example, on 3.14t this grows RSS by about 500 bytes per iteration, which is about 1.4 GB for 3M iterations in a single call:

def f(n: int) -> int:
    cnt = 0
    for i in range(n):
        try:
            raise TypeError("x")
        except TypeError:
            cnt += 1
    return cnt

This PR caches the code objects in a small direct-mapped table keyed by filename, function name and line, guarded by a mutex on free-threaded builds. That fixes the memory growth. It also makes raising exceptions in native code faster on every build, because creating a code object is relatively expensive: the loop above runs about 3.6x faster on 3.14t and 2.2x faster on 3.14.

Added a run test that checks the traceback entries are still correct, that code objects are reused, and that a raise/catch loop doesn't create code-object garbage.