gh-149110: Fix race in _PyFrame_IsIncomplete for FRAME_OWNED_BY_FRAME…

[BHUVANSH855]

_PyFrame_IsIncomplete was checking instr_ptr against a TLBC slot on
frames already owned by a frame object, which races with take_ownership()
when another thread walks frame.f_back (e.g. via pdb).

Fix:

• Short-circuit _PyFrame_IsIncomplete for FRAME_OWNED_BY_FRAME_OBJECT
  frames — take_ownership() already guarantees they're complete
• Fix publication order in take_ownership() so owner is set before
  f->f_frame is visible to concurrent readers
• Use atomic release/acquire for co_tlbc->entries[idx] writes/reads

Added a regression test in test_free_threading/test_frame.py.

Fixes issue #149110

• Issue: Assertion failure: !_PyFrame_IsIncomplete(frame->f_frame) #149110

[markshannon]

Does the C memory model guarantee this?
Why can't another thread see f->f_frame = new_frame before new_frame->owner = FRAME_OWNED_BY_FRAME_OBJECT?

[BHUVANSH855]

You're right, plain C stores provide no ordering guarantee. Fixed in the follow-up commit: take_ownership() now uses _Py_atomic_store_ptr_release() to publish f->f_frame, and PyFrame_GetBack() uses _Py_atomic_load_ptr_acquire() to read it. This establishes the required C11 happens-before relationship — any reader that observes the new f_frame pointer is guaranteed to also observe new_frame->owner == FRAME_OWNED_BY_FRAME_OBJECT.