gh-159312: Resolve CPU attribution from collected stacks

[pablogsal]

CPU mode can attribute samples to a suspended stack when another thread state uses the same native thread. Use the binding status and stacks already collected to prefer a bound, nonempty stack. Preserve the existing attribution when no such stack is present, including native work after a callback deletes or empties its temporary state.

Only samples containing an unbound stack run the local filter. It uses a temporary set and two passes over the collected results, with no additional remote reads or OS thread-status queries. Ordinary bound-state samples add only local checks. The regression uses the existing test helper.

External-inspection and profiling tests pass on regular and free-threaded debug builds, with 547 tests on each; the regression also passes reference-leak checks. Linux syscall counts match the original code for ordinary, duplicate-state, and callback-return samples, with frame caching on and off. In cached ALL-mode debug measurements, median sample times were 6.14 to 6.09 microseconds for ordinary sampling, 19.85 to 20.31 for duplicate states, and 29.00 to 29.56 after a callback. The unusual cases still incur local filtering work.

• Issue: profiling.sampling CPU mode may attribute on-CPU samples to detached tstates when sharing native_thread_id #159312